• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2023 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/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "src/core/CL/CLUtils.h"
29 #include "src/core/CL/CLValidate.h"
30 #include "src/core/experimental/PostOpUtils.h"
31 #include "src/core/helpers/AutoConfiguration.h"
32 #include "src/core/helpers/WindowHelpers.h"
33 #include "src/core/utils/helpers/float_ops.h"
34 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
35 #include "support/Cast.h"
36 #include "support/StringSupport.h"
37 
38 namespace arm_compute
39 {
40 namespace opencl
41 {
42 namespace kernels
43 {
44 namespace
45 {
46 using ElementsProcessed = Steps;
47 
48 const auto post_op_utils = experimental::PostOpCLKernelUtils(
49 {
50     //  PostOp sequence                   -> {Kernel Postfix, PostOp Slots}
51     { {}, { "", {} } },
52     { { experimental::PostOpType::Activation }, { "", { 1 } } },
53 
54     { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } },
55     { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } },
56 
57     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } },
58     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } },
59 
60     { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
61     { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
62 
63     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } },
64     { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }
65 });
66 
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)67 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
68                           const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
69 {
70     ARM_COMPUTE_UNUSED(alpha);
71     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
72     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src0);
73     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32);
74     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
75     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
76     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
77     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");
78     ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16 || rhs_info.k0 < 2);
79     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");
80     ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16 || rhs_info.n0 < 2);
81     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");
82     ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr)
83                                     && (!gemm_info.broadcast_bias),
84                                     "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
85     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
86     ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info));
87     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");
88 
89     const unsigned int m = gemm_info.m;
90     const unsigned int n = gemm_info.n;
91     const unsigned int k = gemm_info.k;
92 
93     TensorShape tensor_shape1{ src1->tensor_shape() };
94     tensor_shape1.set(0, n);
95     tensor_shape1.set(1, k);
96 
97     if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
98     {
99         const unsigned int src2_dim0 = src2->dimension(0);
100         const unsigned int src2_dim1 = src2->dimension(1);
101 
102         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src0);
103         if(gemm_info.broadcast_bias)
104         {
105             ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
106         }
107         else
108         {
109             ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
110         }
111     }
112 
113     const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
114 
115     const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info));
116 
117     ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k);
118     if(gemm_info.reinterpret_input_as_3d)
119     {
120         ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m);
121     }
122     else
123     {
124         ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m);
125     }
126     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
127 
128     if(dst->total_size() != 0)
129     {
130         const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
131         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
132         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
133         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");
134     }
135 
136     return Status{};
137 }
138 
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)139 Window validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
140                                      const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
141 {
142     ARM_COMPUTE_UNUSED(src0, src1, src2);
143     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
144     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
145     bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
146     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
147 
148     // In case both input and dst have to be reinterpreted as 3D tensors,
149     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
150     // This approach should only be used when the input/dst tensors have pad on the y direction
151     if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y)
152     {
153         reinterpret_output_as_3d = false;
154     }
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     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
176     Window             collapsed             = win.collapse(win, dimension_to_collapse);
177 
178     return collapsed;
179 }
180 } // namespace
181 
ClGemmMatrixMultiplyReshapedOnlyRhsKernel()182 ClGemmMatrixMultiplyReshapedOnlyRhsKernel::ClGemmMatrixMultiplyReshapedOnlyRhsKernel()
183 {
184     _type = CLKernelType::GEMM;
185 }
186 
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)187 void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext &compile_context,
188                                                           const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta,
189                                                           const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
190 {
191     ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
192 
193     // dst tensor auto initialization if not yet initialized
194     auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
195 
196     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
197 
198     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
199     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
200     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
201     _add_bias                 = src2 != nullptr;
202     _export_to_cl_image       = rhs_info.export_to_cl_image;
203     _has_pad_y                = gemm_info.has_pad_y;
204     _num_post_op_args         = gemm_info.post_ops.total_num_arguments();
205 
206     auto padding_info = get_padding_info({ src0, src1, src2, dst });
207 
208     // In case both input and dst have to be reinterpreted as 3D tensors,
209     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
210     if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y)
211     {
212         _reinterpret_input_as_3d  = false;
213         _reinterpret_output_as_3d = false;
214     }
215 
216     // Check if we need to slide the matrix B
217     const unsigned int num_dimensions_src0 = src0->num_dimensions();
218     _slide_matrix_b                        = (src1->num_dimensions() >= num_dimensions_src0);
219 
220     ElementsProcessed num_elements_processed{};
221 
222     // Configure kernel window
223     Window win = validate_and_configure_window(src0->clone().get(), src1->clone().get(), (src2 != nullptr) ? src2->clone().get() : nullptr, dst->clone().get(), lhs_info, rhs_info, gemm_info,
224                                                num_elements_processed);
225     ICLKernel::configure_internal(win);
226 
227     // If _reinterpret_input_as_3d = reinterpret_output_as_3d = true,
228     // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
229     // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m
230     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
231 
232     // These variables are used only if gemm_info.has_pad_y == true
233     const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1);
234     const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2);
235 
236     // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
237     // NOTE: This might have implications on heuristics and performance
238     const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
239 
240     // 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.
241     const unsigned int partial_store_m0 = internal_m % internal_m0;
242     const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
243     _m                                  = internal_m;
244     _n                                  = gemm_info.n;
245     _k                                  = gemm_info.k;
246     // Create build options
247     CLBuildOptions build_opts;
248     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
249     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
250     build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
251     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
252     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
253     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
254     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
255     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
256     build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
257     build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(src1->dimension(1)));
258     build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
259     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
260     build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
261     build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
262     build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
263     build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
264     if(_has_pad_y)
265     {
266         build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
267         build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
268         build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
269         build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
270     }
271     // If post_ops are used, then we disable the use of gemm_info.activation_info
272     if(gemm_info.post_ops.size() > 0)
273     {
274         post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops);
275     }
276     else
277     {
278         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
279         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
280         build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
281     }
282 
283     std::string kernel_name("gemm_mm_reshaped_only_rhs_");
284     kernel_name += rhs_info.transpose ? "t" : "nt";
285     kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
286     post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops);
287 
288     // A macro guard to compile ONLY the kernel of interest
289     build_opts.add_option("-D" + upper_string(kernel_name));
290 
291     // Create kernel
292     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
293 
294     // Set config_id for enabling LWS tuning
295     _config_id = kernel_name;
296     _config_id += "_";
297     _config_id += (_has_pad_y ? "" : "no_pad_y_");
298     _config_id += (_add_bias ? "add_bias_" : "");
299     _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
300     _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
301     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
302     _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
303     _config_id += lower_string(string_from_data_type(src0->data_type()));
304     _config_id += "_";
305     _config_id += support::cpp11::to_string(dst->dimension(1));
306     _config_id += "_";
307     _config_id += support::cpp11::to_string(dst->dimension(0));
308     _config_id += "_";
309     _config_id += support::cpp11::to_string(gemm_info.k);
310     _config_id += "_";
311     _config_id += support::cpp11::to_string(dst->dimension(2));
312     _config_id += "_";
313     _config_id += support::cpp11::to_string(lhs_info.m0);
314     _config_id += "_";
315     _config_id += support::cpp11::to_string(rhs_info.n0);
316     _config_id += "_";
317     _config_id += support::cpp11::to_string(rhs_info.k0);
318     _config_id += "_";
319     _config_id += support::cpp11::to_string(rhs_info.h0);
320     _config_id += "_";
321     _config_id += support::cpp11::to_string(rhs_info.interleave);
322 
323     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
324 }
325 
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)326 Status ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
327                                                            const GEMMLHSMatrixInfo &lhs_info,
328                                                            const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
329 {
330     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
331     return Status{};
332 }
333 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)334 void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
335 {
336     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
337     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
338 
339     const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
340     const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
341     const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
342     auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
343 
344     ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
345     ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
346 
347     if(src1->info()->num_dimensions() < 3)
348     {
349         // The stride_z for matrix B must be zero if we do not slice
350         ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
351     }
352 
353     const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u;
354     const size_t rhs_idx_batch_size = 2u;
355     const size_t bia_idx_batch_size = 2u;
356     const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y ? 3u : 2u;
357 
358     Window slice          = window.first_slice_window_3D();
359     Window slice_matrix_b = slice;
360 
361     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
362     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
363 
364     // Get cross plane pads
365     const unsigned int total_cross_plane_pad_lhs = src0->info()->padding().top + src0->info()->padding().bottom;
366     const unsigned int total_cross_plane_pad_out = dst->info()->padding().top + dst->info()->padding().bottom;
367 
368     // 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
369     ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0)));
370 
371     cl::Image2D src1_image2d;
372 
373     if(_export_to_cl_image)
374     {
375         const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2));
376         const size_t      image_row_pitch = src1->info()->strides_in_bytes()[1];
377 
378         src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly);
379     }
380 
381     do
382     {
383         Window slice_b = slice;
384         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
385         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
386         if(!_slide_matrix_b)
387         {
388             slice_b = slice_matrix_b;
389         }
390 
391         unsigned int idx = 0;
392 
393         // LHS buffer
394         add_2D_tensor_argument(idx, src0, slice);
395 
396         // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
397         if(_export_to_cl_image)
398         {
399             _kernel.setArg(idx++, src1_image2d);
400         }
401         else
402         {
403             add_2D_tensor_argument(idx, src1, slice_b);
404         }
405 
406         // Bias buffer (_add_bias == true)
407         add_2D_tensor_argument_if(_add_bias, idx, src2, slice);
408 
409         // dst buffer
410         add_2D_tensor_argument(idx, dst, slice);
411 
412         // post op argument buffers
413         for(size_t i = 0; i < _num_post_op_args; ++i)
414         {
415             const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
416             add_2D_tensor_argument(idx, post_op_arg, slice);
417         }
418 
419         // LHS stride_z
420         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[lhs_idx_batch_size]));
421 
422         // RHS stride_z (not used if _export_to_cl_image == true)
423         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[rhs_idx_batch_size]));
424 
425         // Bias stride_z (if _add_bias == true)
426         if(_add_bias)
427         {
428             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[bia_idx_batch_size]));
429         }
430 
431         // dst stride_z
432         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size]));
433         // post op argument stride_z
434         for(size_t i = 0; i < _num_post_op_args; ++i)
435         {
436             const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
437             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
438         }
439 
440         // Cross-plan padding (if _reinterpret_input_as_3d = true)
441         if(_reinterpret_input_as_3d && _has_pad_y)
442         {
443             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs));
444         }
445 
446         // Cross-plan padding (if reinterpret_output_as_3d = true)
447         if(_reinterpret_output_as_3d && _has_pad_y)
448         {
449             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out));
450         }
451 
452         // Pass m, n and k at runtime as signed ints, to ensure results of any subractions they could be operand in, would still be signed.
453         _kernel.setArg<cl_int>(idx++, _m);
454         _kernel.setArg<cl_int>(idx++, _n);
455         _kernel.setArg<cl_int>(idx++, _k);
456 
457         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
458     }
459     while(window.slide_window_slice_3D(slice));
460 }
461 } // namespace kernels
462 } // namespace opencl
463 } // namespace arm_compute
464