• 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 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.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/Utils.h"
33 #include "arm_compute/core/Validate.h"
34 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
35 #include "src/core/AccessWindowStatic.h"
36 #include "src/core/CL/CLUtils.h"
37 #include "src/core/CL/CLValidate.h"
38 #include "src/core/CL/gemm/CLGEMMHelpers.h"
39 #include "src/core/helpers/AutoConfiguration.h"
40 #include "src/core/helpers/WindowHelpers.h"
41 #include "src/core/utils/helpers/float_ops.h"
42 #include "support/StringSupport.h"
43 
44 #include <cstddef>
45 #include <cstdint>
46 #include <tuple>
47 
48 using namespace arm_compute;
49 using namespace arm_compute::misc::shape_calculator;
50 
51 namespace arm_compute
52 {
53 class Coordinates;
54 } // namespace arm_compute
55 
56 namespace
57 {
58 using ElementsProcessed = Steps;
59 
validate_arguments(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)60 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
61                           const GEMMRHSMatrixInfo &rhs_info,
62                           const GEMMKernelInfo    &gemm_info)
63 {
64     ARM_COMPUTE_UNUSED(alpha);
65     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
66     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0);
67     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
68     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
69     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
70     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
71     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
72     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose == rhs_info.transpose);
73     ARM_COMPUTE_RETURN_ERROR_ON_MSG(((lhs_info.k0 & (lhs_info.k0 - 1)) && lhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
74     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
75     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
76     ARM_COMPUTE_RETURN_ERROR_ON_MSG((lhs_info.transpose) && ((lhs_info.m0 & (lhs_info.m0 - 1)) && lhs_info.m0 != 3), "Only 2,3,4,8,16 are supported for m0");
77     ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.transpose) && ((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
78     ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (input2 != nullptr)
79                                     && (!gemm_info.broadcast_bias),
80                                     "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");
81     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision && (input0->data_type() == DataType::F32), "Mixed precision only supported for F16 data type");
82     ARM_COMPUTE_RETURN_ON_ERROR(cl_gemm::validate_image2d_support_on_rhs(*input1, rhs_info));
83 
84     const unsigned int m = gemm_info.m;
85     const unsigned int n = gemm_info.n;
86     const unsigned int k = gemm_info.k;
87 
88     TensorShape tensor_shape0{ input0->tensor_shape() };
89     tensor_shape0.set(0, k);
90     tensor_shape0.set(1, m);
91 
92     TensorShape tensor_shape1{ input1->tensor_shape() };
93     tensor_shape1.set(0, n);
94     tensor_shape1.set(1, k);
95 
96     if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
97     {
98         const unsigned int input2_dim0 = input2->dimension(0);
99         const unsigned int input2_dim1 = input2->dimension(1);
100 
101         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input1);
102         if(gemm_info.broadcast_bias)
103         {
104             ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
105         }
106         else
107         {
108             ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
109         }
110     }
111 
112     const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
113     const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
114 
115     const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
116     const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
117 
118     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
119     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
120 
121     if(output->total_size() != 0)
122     {
123         const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
124         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
125         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
126     }
127 
128     return Status{};
129 }
130 
validate_and_configure_window(ITensorInfo * input0,ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info,ElementsProcessed & num_elements_processed)131 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
132                                                         const GEMMRHSMatrixInfo &rhs_info,
133                                                         const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
134 {
135     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
136     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
137     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
138 
139     Window win{};
140     Window win_out{};
141     bool   window_changed = false;
142 
143     // Output tensor auto initialization if not yet initialized
144     auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));
145 
146     TensorInfo tmp_info(*output);
147 
148     if(reinterpret_output_as_3d)
149     {
150         // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
151         // the window needs to be constructed on the 2D collapsed version of the tensor
152         TensorShape tmp_shape(output->tensor_shape());
153         tmp_shape.collapse(2U, 1U);
154         tmp_info.set_tensor_shape(tmp_shape);
155     }
156 
157     // Configure kernel window
158     num_elems_processed_per_iteration_x = rhs_info.n0;
159     num_elems_processed_per_iteration_y = lhs_info.m0;
160 
161     win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
162     win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
163 
164     AccessWindowStatic input0_access(input0, 0, 0,
165                                      input0->dimension(0),
166                                      input0->dimension(1));
167     AccessWindowStatic input1_access(input1, 0, 0,
168                                      input1->dimension(0),
169                                      input1->dimension(1));
170     AccessWindowStatic output_access(output, 0, 0,
171                                      output->dimension(0),
172                                      output->dimension(1));
173 
174     if(input2 != nullptr)
175     {
176         const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
177 
178         const int bias_processed_per_iteration_y = gemm_info.broadcast_bias ? 1 : num_elems_processed_per_iteration_y;
179 
180         AccessWindowStatic input2_access(input2, 0, 0,
181                                          ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
182                                          ceil_to_multiple(input2->dimension(1), bias_processed_per_iteration_y));
183 
184         window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
185                          update_window_and_padding(win_out, output_access);                             // window used to update the padding requirements of output tensor
186     }
187     else
188     {
189         window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
190                          update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
191     }
192 
193     output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
194 
195     // Collapse along the Z direction
196     // This collapse needs to be here in order to tune the Z dimension of LWS
197     Window             collapsed             = win;
198     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
199     collapsed                                = win.collapse(win, dimension_to_collapse);
200 
201     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
202     return std::make_pair(err, collapsed);
203 }
204 } // namespace
205 
CLGEMMMatrixMultiplyReshapedKernel()206 CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel()
207     : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _use_dummy_work_items(false), _add_bias(false),
208       _broadcast_bias(false), _export_to_cl_image(false), _k(1)
209 {
210 }
211 
configure(const ICLTensor * input0,const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)212 void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
213                                                    const GEMMLHSMatrixInfo &lhs_info,
214                                                    const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
215 {
216     configure(CLKernelLibrary::get().get_compile_context(), input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info);
217 }
218 
configure(const CLCompileContext & compile_context,const ICLTensor * input0,const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)219 void CLGEMMMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha,
220                                                    float                    beta,
221                                                    const GEMMLHSMatrixInfo &lhs_info,
222                                                    const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
223 {
224     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
225 
226     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));
227 
228     auto padding_info         = get_padding_info({ input0, output });
229     _input0                   = input0;
230     _input1                   = input1;
231     _input2                   = helpers::float_ops::is_zero(beta) ? nullptr : input2;
232     _output                   = output;
233     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
234     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
235     _add_bias                 = _input2 != nullptr;
236     _broadcast_bias           = gemm_info.broadcast_bias;
237     _export_to_cl_image       = rhs_info.export_to_cl_image;
238     _k                        = gemm_info.k;
239 
240     // Check if we need to slide the matrix B
241     const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
242     _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);
243 
244     ElementsProcessed num_elements_processed{};
245 
246     // Configure kernel window
247     auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
248     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
249     ICLKernel::configure_internal(win_config.second);
250 
251     const bool     enable_mixed_precision = gemm_info.fp_mixed_precision;
252     const DataType data_type              = input0->info()->data_type();
253 
254     // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
255     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
256 
257     const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
258     const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
259 
260     // Create build options
261     CLBuildOptions build_opts;
262     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
263     build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
264     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
265     build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
266     build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
267     build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
268     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
269     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
270     build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
271     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
272     build_opts.add_option_if(lhs_info.transpose, "-DLHS_TRANSPOSE");
273     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
274     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
275     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
276     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
277     build_opts.add_option_if(enable_mixed_precision, "-DMIXED_PRECISION");
278     build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
279     build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(input1->info()->dimension(1)));
280     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
281     build_opts.add_option("-DDATA_TYPE_ACCUMULATOR=" + (enable_mixed_precision ? get_cl_type_from_data_type(DataType::F32) : get_cl_type_from_data_type(data_type)));
282     build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m));
283     build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
284     build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
285     build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
286     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
287     build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
288     build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
289     build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
290     build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
291     build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
292 
293     std::string kernel_name("gemm_mm_reshaped_");
294     kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
295     kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
296     kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
297 
298     // Create kernel
299     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
300 
301     // Set config_id for enabling LWS tuning
302     _config_id = kernel_name;
303     _config_id += "_";
304     _config_id += (_add_bias ? "add_bias_" : "");
305     _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
306     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
307     _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
308     _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
309     _config_id += "_";
310     _config_id += (enable_mixed_precision ? "mixed_precision_" : "");
311     _config_id += support::cpp11::to_string(output->info()->dimension(1));
312     _config_id += "_";
313     _config_id += support::cpp11::to_string(output->info()->dimension(0));
314     _config_id += "_";
315     _config_id += support::cpp11::to_string(gemm_info.k);
316     _config_id += "_";
317     _config_id += support::cpp11::to_string(output->info()->dimension(2));
318     _config_id += "_";
319     _config_id += support::cpp11::to_string(lhs_info.m0);
320     _config_id += "_";
321     _config_id += support::cpp11::to_string(rhs_info.n0);
322     _config_id += "_";
323     _config_id += support::cpp11::to_string(lhs_info.k0);
324     _config_id += "_";
325     _config_id += support::cpp11::to_string(lhs_info.v0);
326     _config_id += "_";
327     _config_id += support::cpp11::to_string(rhs_info.h0);
328     _config_id += "_";
329     _config_id += support::cpp11::to_string(lhs_info.interleave);
330     _config_id += "_";
331     _config_id += support::cpp11::to_string(rhs_info.interleave);
332 
333     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
334 }
335 
validate(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)336 Status CLGEMMMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
337                                                     const GEMMLHSMatrixInfo &lhs_info,
338                                                     const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
339 {
340     ElementsProcessed num_elements_processed{};
341     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
342     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
343                                                               input1->clone().get(),
344                                                               input2 != nullptr ? input2->clone().get() : nullptr,
345                                                               output->clone().get(),
346                                                               lhs_info,
347                                                               rhs_info,
348                                                               gemm_info,
349                                                               num_elements_processed)
350                                 .first);
351 
352     return Status{};
353 }
354 
run(const Window & window,cl::CommandQueue & queue)355 void CLGEMMMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQueue &queue)
356 {
357     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
358     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
359 
360     if(_input1->info()->num_dimensions() < 3)
361     {
362         // The stride_z for matrix B must be zero if we do not slice
363         ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
364     }
365 
366     Window slice          = window.first_slice_window_3D();
367     Window slice_matrix_b = slice;
368 
369     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
370     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
371 
372     const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
373 
374     cl::Image2D input1_image2d;
375 
376     if(_export_to_cl_image)
377     {
378         const TensorShape shape2d(_input1->info()->dimension(0) / 4, _input1->info()->dimension(1) * _input1->info()->dimension(2));
379         const size_t      image_row_pitch = _input1->info()->strides_in_bytes()[1];
380 
381         input1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), _input1->cl_buffer(), shape2d, _input1->info()->data_type(), image_row_pitch);
382     }
383 
384     do
385     {
386         Window slice_b = slice;
387         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
388         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
389         if(!_slide_matrix_b)
390         {
391             slice_b = slice_matrix_b;
392         }
393 
394         unsigned int idx = 0;
395 
396         // LHS buffer
397         add_2D_tensor_argument(idx, _input0, slice);
398 
399         // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
400         if(_export_to_cl_image)
401         {
402             _kernel.setArg(idx++, input1_image2d);
403         }
404         else
405         {
406             add_2D_tensor_argument(idx, _input1, slice_b);
407         }
408 
409         // Bias buffer (_add_bias == true)
410         add_2D_tensor_argument_if(_add_bias, idx, _input2, slice);
411 
412         // Output buffer
413         add_2D_tensor_argument(idx, _output, slice);
414 
415         // K dimension (not used if _export_to_cl_image == true)
416         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
417 
418         // LHS stride_z
419         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
420 
421         // RHS stride_z (not used if _export_to_cl_image == true)
422         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
423 
424         // Bias stride_z (if _add_bias == true)
425         if(_add_bias)
426         {
427             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
428         }
429 
430         // Output stride_z
431         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
432 
433         // Cross-plan padding (if _reinterpret_output_as_3d = true)
434         if(_reinterpret_output_as_3d)
435         {
436             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad));
437         }
438 
439         // Dispatch kernel
440         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
441     }
442     while(window.slide_window_slice_3D(slice));
443 }
444