• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2020 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/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Utils.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "src/core/AccessWindowStatic.h"
32 #include "src/core/CL/CLUtils.h"
33 #include "src/core/CL/CLValidate.h"
34 #include "src/core/CL/gemm/CLGEMMHelpers.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 #include "src/core/utils/helpers/float_ops.h"
38 #include "support/StringSupport.h"
39 
40 #include <tuple>
41 
42 using namespace arm_compute::misc::shape_calculator;
43 
44 namespace arm_compute
45 {
46 namespace
47 {
48 using ElementsProcessed = Steps;
49 
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)50 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
51                           const GEMMRHSMatrixInfo &rhs_info,
52                           const GEMMKernelInfo    &gemm_info)
53 {
54     ARM_COMPUTE_UNUSED(alpha);
55     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
56     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0);
57     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
58     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
59     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
60     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
61     ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_info.m0 < 1 || lhs_info.m0 > 8, "Only 1,2,3,4,5,6,7,8 are supported for m0");
62     ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16 || rhs_info.k0 < 2);
63     ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
64     ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16 || rhs_info.n0 < 2);
65     ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
66     ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (input2 != nullptr)
67                                     && (!gemm_info.broadcast_bias),
68                                     "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");
69     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
70     ARM_COMPUTE_RETURN_ON_ERROR(cl_gemm::validate_image2d_support_on_rhs(*input1, rhs_info));
71 
72     const unsigned int m = gemm_info.m;
73     const unsigned int n = gemm_info.n;
74     const unsigned int k = gemm_info.k;
75 
76     TensorShape tensor_shape1{ input1->tensor_shape() };
77     tensor_shape1.set(0, n);
78     tensor_shape1.set(1, k);
79 
80     if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
81     {
82         const unsigned int input2_dim0 = input2->dimension(0);
83         const unsigned int input2_dim1 = input2->dimension(1);
84 
85         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input0);
86         if(gemm_info.broadcast_bias)
87         {
88             ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
89         }
90         else
91         {
92             ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
93         }
94     }
95 
96     const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
97 
98     const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
99 
100     ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != k);
101     if(gemm_info.reinterpret_input_as_3d)
102     {
103         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != m);
104     }
105     else
106     {
107         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != m);
108     }
109     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
110 
111     if(output->total_size() != 0)
112     {
113         const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
114         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
115         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
116     }
117 
118     return Status{};
119 }
120 
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)121 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
122                                                         const GEMMRHSMatrixInfo &rhs_info,
123                                                         const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
124 {
125     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
126     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
127     bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
128     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
129 
130     Window win{};
131     Window win_out{};
132     bool   window_changed = false;
133 
134     // In case both input and output have to be reinterpreted as 3D tensors,
135     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
136     // This approach should only be used when the input/output tensors have pad on the y direction
137     if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y)
138     {
139         reinterpret_output_as_3d = false;
140     }
141 
142     // Output tensor auto initialization if not yet initialized
143     auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));
144 
145     TensorInfo tmp_info(*output);
146 
147     if(reinterpret_output_as_3d)
148     {
149         // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
150         // the window needs to be constructed on the 2D collapsed version of the tensor
151         TensorShape tmp_shape(output->tensor_shape());
152         tmp_shape.collapse(2U, 1U);
153         tmp_info.set_tensor_shape(tmp_shape);
154     }
155 
156     // Configure kernel window
157     num_elems_processed_per_iteration_x = rhs_info.n0;
158     num_elems_processed_per_iteration_y = lhs_info.m0;
159 
160     win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
161     win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
162 
163     if(input2 != nullptr)
164     {
165         const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
166 
167         AccessWindowStatic input2_access(input2, 0, 0,
168                                          ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
169                                          input2->dimension(1));
170 
171         window_changed = update_window_and_padding(win, input2_access);
172     }
173 
174     // Collapse along the Z direction
175     // This collapse needs to be here in order to tune the Z dimension of LWS
176     Window             collapsed             = win;
177     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
178     collapsed                                = win.collapse(win, dimension_to_collapse);
179 
180     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
181     return std::make_pair(err, collapsed);
182 }
183 } // namespace
184 
CLGEMMMatrixMultiplyReshapedOnlyRHSKernel()185 CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMMatrixMultiplyReshapedOnlyRHSKernel()
186     : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false),
187       _add_bias(false), _broadcast_bias(false), _export_to_cl_image(false), _has_pad_y(false)
188 {
189 }
190 
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)191 void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
192                                                           const GEMMLHSMatrixInfo &lhs_info,
193                                                           const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
194 {
195     configure(CLKernelLibrary::get().get_compile_context(), input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info);
196 }
197 
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)198 void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output,
199                                                           float                    alpha,
200                                                           float                    beta,
201                                                           const GEMMLHSMatrixInfo &lhs_info,
202                                                           const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
203 {
204     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
205 
206     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));
207 
208     _input0                   = input0;
209     _input1                   = input1;
210     _input2                   = helpers::float_ops::is_zero(beta) ? nullptr : input2;
211     _output                   = output;
212     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
213     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
214     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
215     _add_bias                 = _input2 != nullptr;
216     _broadcast_bias           = gemm_info.broadcast_bias;
217     _export_to_cl_image       = rhs_info.export_to_cl_image;
218     _has_pad_y                = gemm_info.has_pad_y;
219 
220     auto padding_info = get_padding_info({ input0, input1, output });
221 
222     // In case both input and output have to be reinterpreted as 3D tensors,
223     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
224     if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y)
225     {
226         _reinterpret_input_as_3d  = false;
227         _reinterpret_output_as_3d = false;
228     }
229 
230     // Check if we need to slide the matrix B
231     const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
232     _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);
233 
234     ElementsProcessed num_elements_processed{};
235 
236     // Configure kernel window
237     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);
238     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
239     ICLKernel::configure_internal(win_config.second);
240 
241     // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
242     // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
243     // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m
244     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
245 
246     // These variables are used only if gemm_info.has_pad_y == true
247     const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1);
248     const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2);
249 
250     // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
251     // NOTE: This might have implications on heuristics and performance
252     const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
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 partial_store_m0 = internal_m % internal_m0;
256     const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
257 
258     // Create build options
259     CLBuildOptions build_opts;
260     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
261     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
262     build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
263     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
264     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
265     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
266     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
267     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
268     build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
269     build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(input1->info()->dimension(1)));
270     build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m));
271     build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
272     build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
273     build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
274     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
275     build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
276     build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
277     build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
278     build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
279     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
280     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
281     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
282     if(_has_pad_y)
283     {
284         build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
285         build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
286         build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
287         build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
288     }
289 
290     std::string kernel_name("gemm_mm_reshaped_only_rhs_");
291     kernel_name += rhs_info.transpose ? "t" : "nt";
292     kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
293 
294     // Create kernel
295     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
296 
297     // Set config_id for enabling LWS tuning
298     _config_id = kernel_name;
299     _config_id += "_";
300     _config_id += (_has_pad_y ? "" : "no_pad_y_");
301     _config_id += (_add_bias ? "add_bias_" : "");
302     _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
303     _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
304     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
305     _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
306     _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
307     _config_id += "_";
308     _config_id += support::cpp11::to_string(output->info()->dimension(1));
309     _config_id += "_";
310     _config_id += support::cpp11::to_string(output->info()->dimension(0));
311     _config_id += "_";
312     _config_id += support::cpp11::to_string(gemm_info.k);
313     _config_id += "_";
314     _config_id += support::cpp11::to_string(output->info()->dimension(2));
315     _config_id += "_";
316     _config_id += support::cpp11::to_string(lhs_info.m0);
317     _config_id += "_";
318     _config_id += support::cpp11::to_string(rhs_info.n0);
319     _config_id += "_";
320     _config_id += support::cpp11::to_string(rhs_info.k0);
321     _config_id += "_";
322     _config_id += support::cpp11::to_string(rhs_info.h0);
323     _config_id += "_";
324     _config_id += support::cpp11::to_string(rhs_info.interleave);
325 
326     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
327 }
328 
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)329 Status CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
330                                                            const GEMMLHSMatrixInfo &lhs_info,
331                                                            const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
332 {
333     ElementsProcessed num_elements_processed{};
334     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
335     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
336                                                               input1->clone().get(),
337                                                               input2 != nullptr ? input2->clone().get() : nullptr,
338                                                               output->clone().get(),
339                                                               lhs_info,
340                                                               rhs_info,
341                                                               gemm_info,
342                                                               num_elements_processed)
343                                 .first);
344 
345     return Status{};
346 }
347 
run(const Window & window,cl::CommandQueue & queue)348 void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::CommandQueue &queue)
349 {
350     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
351     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
352 
353     if(_input1->info()->num_dimensions() < 3)
354     {
355         // The stride_z for matrix B must be zero if we do not slice
356         ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
357     }
358 
359     const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u;
360     const size_t rhs_idx_batch_size = 2u;
361     const size_t bia_idx_batch_size = 2u;
362     const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y ? 3u : 2u;
363 
364     Window slice          = window.first_slice_window_3D();
365     Window slice_matrix_b = slice;
366 
367     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
368     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
369 
370     // Get cross plane pads
371     const unsigned int total_cross_plane_pad_lhs = _input0->info()->padding().top + _input0->info()->padding().bottom;
372     const unsigned int total_cross_plane_pad_out = _output->info()->padding().top + _output->info()->padding().bottom;
373 
374     // The execution should fail if we try to run with has_pad_y = false but we have padding in either the LHS or DST tensor
375     ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0)));
376 
377     cl::Image2D input1_image2d;
378 
379     if(_export_to_cl_image)
380     {
381         const TensorShape shape2d(_input1->info()->dimension(0) / 4, _input1->info()->dimension(1) * _input1->info()->dimension(2));
382         const size_t      image_row_pitch = _input1->info()->strides_in_bytes()[1];
383 
384         input1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), _input1->cl_buffer(), shape2d, _input1->info()->data_type(), image_row_pitch);
385     }
386 
387     do
388     {
389         Window slice_b = slice;
390         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
391         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
392         if(!_slide_matrix_b)
393         {
394             slice_b = slice_matrix_b;
395         }
396 
397         unsigned int idx = 0;
398 
399         // LHS buffer
400         add_2D_tensor_argument(idx, _input0, slice);
401 
402         // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
403         if(_export_to_cl_image)
404         {
405             _kernel.setArg(idx++, input1_image2d);
406         }
407         else
408         {
409             add_2D_tensor_argument(idx, _input1, slice_b);
410         }
411 
412         // Bias buffer (_add_bias == true)
413         add_2D_tensor_argument_if(_add_bias, idx, _input2, slice);
414 
415         // Output buffer
416         add_2D_tensor_argument(idx, _output, slice);
417 
418         // LHS stride_z
419         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[lhs_idx_batch_size]));
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()[rhs_idx_batch_size]));
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()[bia_idx_batch_size]));
428         }
429 
430         // Output stride_z
431         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[out_idx_batch_size]));
432 
433         // Cross-plan padding (if _reinterpret_input_as_3d = true)
434         if(_reinterpret_input_as_3d && _has_pad_y)
435         {
436             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs));
437         }
438 
439         // Cross-plan padding (if _reinterpret_output_as_3d = true)
440         if(_reinterpret_output_as_3d && _has_pad_y)
441         {
442             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out));
443         }
444 
445         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
446     }
447     while(window.slide_window_slice_3D(slice));
448 }
449 } // namespace arm_compute
450