• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2021, 2023 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/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/CL/CLUtils.h"
36 #include "src/core/CL/CLValidate.h"
37 #include "src/core/experimental/PostOpUtils.h"
38 #include "src/core/helpers/AutoConfiguration.h"
39 #include "src/core/helpers/WindowHelpers.h"
40 #include "src/core/utils/helpers/float_ops.h"
41 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
42 #include "support/Cast.h"
43 #include "support/StringSupport.h"
44 
45 namespace arm_compute
46 {
47 namespace opencl
48 {
49 namespace kernels
50 {
51 namespace
52 {
53 using ElementsProcessed = Steps;
54 
55 const auto post_op_utils = experimental::PostOpCLKernelUtils(
56 {
57     //  PostOp sequence                   -> {Kernel Postfix, PostOp Slots}
58     { {}, { "", {} } },
59     { { experimental::PostOpType::Activation }, { "", { 1 } } },
60 
61     { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } },
62     { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } },
63 
64     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } },
65     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } },
66 
67     { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
68     { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
69 
70     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } },
71     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }
72 });
73 
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)74 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
75                           const GEMMRHSMatrixInfo &rhs_info,
76                           const GEMMKernelInfo    &gemm_info)
77 {
78     ARM_COMPUTE_UNUSED(alpha);
79     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
80     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src0);
81     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32);
82     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
83     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
84     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
85     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
86     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose == rhs_info.transpose);
87     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");
88     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
89     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
90     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");
91     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");
92     ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr)
93                                     && (!gemm_info.broadcast_bias),
94                                     "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
95     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision && (src0->data_type() == DataType::F32), "Mixed precision only supported for F16 data type");
96     ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info));
97     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported");
98 
99     const unsigned int m = gemm_info.m;
100     const unsigned int n = gemm_info.n;
101     const unsigned int k = gemm_info.k;
102 
103     TensorShape tensor_shape0{ src0->tensor_shape() };
104     tensor_shape0.set(0, k);
105     tensor_shape0.set(1, m);
106 
107     TensorShape tensor_shape1{ src1->tensor_shape() };
108     tensor_shape1.set(0, n);
109     tensor_shape1.set(1, k);
110 
111     if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
112     {
113         const unsigned int src2_dim0 = src2->dimension(0);
114         const unsigned int src2_dim1 = src2->dimension(1);
115 
116         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
117         if(gemm_info.broadcast_bias)
118         {
119             ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
120         }
121         else
122         {
123             ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
124         }
125     }
126 
127     const TensorInfo tensor_info0 = src0->clone()->set_tensor_shape(tensor_shape0);
128     const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
129 
130     const TensorInfo tensor_info_reshaped0 = src0->clone()->set_tensor_shape(misc::shape_calculator::compute_lhs_reshaped_shape(tensor_info0, lhs_info));
131     const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info));
132 
133     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src0, &tensor_info_reshaped0);
134     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
135 
136     if(dst->total_size() != 0)
137     {
138         const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
139         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
140         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
141         ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.are_post_op_shapes_compliant(dst, gemm_info.post_ops), "The Post Op shapes are not compliant");
142     }
143 
144     return Status{};
145 }
146 
validate_and_configure_window(ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * src2,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info,ElementsProcessed & num_elements_processed)147 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
148                                                         const GEMMRHSMatrixInfo &rhs_info,
149                                                         const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
150 {
151     ARM_COMPUTE_UNUSED(src0, src1, src2);
152     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
153     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
154     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
155 
156     TensorInfo tmp_info(*dst);
157 
158     if(reinterpret_output_as_3d)
159     {
160         // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
161         // the window needs to be constructed on the 2D collapsed version of the tensor
162         TensorShape tmp_shape(dst->tensor_shape());
163         tmp_shape.collapse(2U, 1U);
164         tmp_info.set_tensor_shape(tmp_shape);
165     }
166 
167     // Configure kernel window
168     num_elems_processed_per_iteration_x = rhs_info.n0;
169     num_elems_processed_per_iteration_y = lhs_info.m0;
170 
171     Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
172 
173     // Collapse along the Z direction
174     // This collapse needs to be here in order to tune the Z dimension of LWS
175     Window             collapsed             = win;
176     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
177     collapsed                                = win.collapse(win, dimension_to_collapse);
178 
179     return std::make_pair(Status{}, collapsed);
180 }
181 } // namespace
182 
ClGemmMatrixMultiplyReshapedKernel()183 ClGemmMatrixMultiplyReshapedKernel::ClGemmMatrixMultiplyReshapedKernel()
184 {
185     _type = CLKernelType::GEMM;
186 }
187 
configure(const CLCompileContext & compile_context,const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)188 void ClGemmMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context,
189                                                    const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta,
190                                                    const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
191 {
192     ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
193 
194     // dst tensor auto initialization if not yet initialized
195     auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
196 
197     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
198 
199     auto padding_info         = get_padding_info({ src0, src1, src2, dst });
200     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
201     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
202     _add_bias                 = src2 != nullptr;
203     _export_to_cl_image       = rhs_info.export_to_cl_image;
204     _num_post_op_args         = gemm_info.post_ops.total_num_arguments();
205 
206     // Check if we need to slide the matrix B
207     const unsigned int num_dimensions_src0 = src0->num_dimensions();
208     _slide_matrix_b                        = (src1->num_dimensions() >= num_dimensions_src0);
209 
210     ElementsProcessed num_elements_processed{};
211 
212     // Configure kernel window
213     auto win_config = validate_and_configure_window(src0->clone().get(),
214                                                     src1->clone().get(),
215                                                     (src2 != nullptr) ? src2->clone().get() : nullptr,
216                                                     dst->clone().get(),
217                                                     lhs_info,
218                                                     rhs_info,
219                                                     gemm_info,
220                                                     num_elements_processed);
221     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
222     ICLKernel::configure_internal(win_config.second);
223 
224     const bool     enable_mixed_precision = gemm_info.fp_mixed_precision;
225     const DataType data_type              = src0->data_type();
226 
227     // 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.
228     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
229 
230     const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
231     const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
232     _m                                  = gemm_info.m;
233     _n                                  = gemm_info.n;
234     _k                                  = gemm_info.k;
235 
236     // Create build options
237     CLBuildOptions build_opts;
238     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
239     build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
240     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
241     build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
242     build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
243     build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
244     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
245     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
246     build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
247     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
248     build_opts.add_option_if(lhs_info.transpose, "-DLHS_TRANSPOSE");
249     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
250     build_opts.add_option_if(enable_mixed_precision, "-DMIXED_PRECISION");
251     build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
252     build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(src1->dimension(1)));
253     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
254     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)));
255     build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
256     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
257     build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
258     build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
259     build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
260     build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
261     build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
262     // If post_ops are used, then we disable the use of gemm_info.activation_info
263     if(gemm_info.post_ops.size() > 0)
264     {
265         post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops);
266     }
267     else
268     {
269         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
270         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
271         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
272     }
273 
274     std::string kernel_name("gemm_mm_reshaped_");
275     kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
276     kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
277     kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
278     post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops);
279 
280     // A macro guard to compile ONLY the kernel of interest
281     build_opts.add_option("-D" + upper_string(kernel_name));
282 
283     // Create kernel
284     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
285 
286     // Set config_id for enabling LWS tuning
287     _config_id = kernel_name;
288     _config_id += "_";
289     _config_id += (_add_bias ? "add_bias_" : "");
290     _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
291     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
292     _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
293     _config_id += lower_string(string_from_data_type(src0->data_type()));
294     _config_id += "_";
295     _config_id += (enable_mixed_precision ? "mixed_precision_" : "");
296     _config_id += support::cpp11::to_string(dst->dimension(1));
297     _config_id += "_";
298     _config_id += support::cpp11::to_string(dst->dimension(0));
299     _config_id += "_";
300     _config_id += support::cpp11::to_string(gemm_info.k);
301     _config_id += "_";
302     _config_id += support::cpp11::to_string(dst->dimension(2));
303     _config_id += "_";
304     _config_id += support::cpp11::to_string(lhs_info.m0);
305     _config_id += "_";
306     _config_id += support::cpp11::to_string(rhs_info.n0);
307     _config_id += "_";
308     _config_id += support::cpp11::to_string(lhs_info.k0);
309     _config_id += "_";
310     _config_id += support::cpp11::to_string(lhs_info.v0);
311     _config_id += "_";
312     _config_id += support::cpp11::to_string(rhs_info.h0);
313     _config_id += "_";
314     _config_id += support::cpp11::to_string(lhs_info.interleave);
315     _config_id += "_";
316     _config_id += support::cpp11::to_string(rhs_info.interleave);
317 
318     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
319 }
320 
validate(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)321 Status ClGemmMatrixMultiplyReshapedKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
322                                                     const GEMMLHSMatrixInfo &lhs_info,
323                                                     const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
324 {
325     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
326     return Status{};
327 }
328 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)329 void ClGemmMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
330 {
331     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
332     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
333 
334     const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
335     const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
336     const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
337     auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
338 
339     ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
340     ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
341 
342     if(src1->info()->num_dimensions() < 3)
343     {
344         // The stride_z for matrix B must be zero if we do not slice
345         ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
346     }
347 
348     Window slice          = window.first_slice_window_3D();
349     Window slice_matrix_b = slice;
350 
351     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
352     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
353 
354     const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
355 
356     cl::Image2D src1_image2d;
357 
358     if(_export_to_cl_image)
359     {
360         const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2));
361         const size_t      image_row_pitch = src1->info()->strides_in_bytes()[1];
362 
363         src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly);
364     }
365 
366     do
367     {
368         Window slice_b = slice;
369         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
370         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
371         if(!_slide_matrix_b)
372         {
373             slice_b = slice_matrix_b;
374         }
375 
376         unsigned int idx = 0;
377 
378         // LHS buffer
379         add_2D_tensor_argument(idx, src0, slice);
380 
381         // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
382         if(_export_to_cl_image)
383         {
384             _kernel.setArg(idx++, src1_image2d);
385         }
386         else
387         {
388             add_2D_tensor_argument(idx, src1, slice_b);
389         }
390 
391         // Bias buffer (_add_bias == true)
392         add_2D_tensor_argument_if(_add_bias, idx, src2, slice);
393 
394         // dst buffer
395         add_2D_tensor_argument(idx, dst, slice);
396 
397         // post op argument buffers
398         for(size_t i = 0; i < _num_post_op_args; ++i)
399         {
400             const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
401             add_2D_tensor_argument(idx, post_op_arg, slice);
402         }
403 
404         // LHS stride_z
405         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
406 
407         // RHS stride_z (not used if _export_to_cl_image == true)
408         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
409 
410         // Bias stride_z (if _add_bias == true)
411         if(_add_bias)
412         {
413             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
414         }
415 
416         // dst stride_z
417         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
418 
419         // post op argument stride_z
420         for(size_t i = 0; i < _num_post_op_args; ++i)
421         {
422             const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
423             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
424         }
425         // Cross-plan padding (if _reinterpret_output_as_3d = true)
426         if(_reinterpret_output_as_3d)
427         {
428             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad));
429         }
430 
431         // Pass m, n and k at runtime
432         _kernel.setArg<cl_int>(idx++, _m);
433         _kernel.setArg<cl_int>(idx++, _n);
434 
435         // K dimension (not used if _export_to_cl_image == true)
436         _kernel.setArg<cl_int>(idx++, _k);
437 
438         // Dispatch kernel
439         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
440     }
441     while(window.slide_window_slice_3D(slice));
442 }
443 } // namespace kernels
444 } // namespace opencl
445 } // namespace arm_compute
446