1 /*
2 * Copyright (c) 2019-2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.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/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 "support/Cast.h"
42 #include "support/StringSupport.h"
43
44 namespace arm_compute
45 {
46 namespace opencl
47 {
48 namespace kernels
49 {
50 namespace
51 {
52 using ElementsProcessed = Steps;
53
54 const auto post_op_utils = experimental::PostOpCLKernelUtils(
55 {
56 // PostOp sequence -> {Kernel Postfix, PostOp Slots}
57 { {}, { "", {} } },
58 { { experimental::PostOpType::Activation }, { "", { 1 } } },
59
60 { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } },
61 { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } },
62
63 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } },
64 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } },
65
66 { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
67 { { experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
68
69 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } },
70 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 1, 2, 3 } } }
71 });
72
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)73 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
74 const GEMMRHSMatrixInfo &rhs_info,
75 const GEMMKernelInfo &gemm_info)
76 {
77 ARM_COMPUTE_UNUSED(alpha);
78 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
79 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F32, DataType::F16);
80 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
81 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
82 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
83 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");
84 ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
85 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
86 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");
87 ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr)
88 && (!gemm_info.broadcast_bias),
89 "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
90 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
91 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native");
92 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");
93
94 const unsigned int m = gemm_info.m;
95 const unsigned int n = gemm_info.n;
96 const unsigned int k = gemm_info.k;
97
98 ARM_COMPUTE_UNUSED(m);
99 ARM_COMPUTE_UNUSED(n);
100 ARM_COMPUTE_UNUSED(k);
101
102 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k);
103 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) != n);
104 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(1) != k);
105 if(gemm_info.reinterpret_input_as_3d)
106 {
107 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m);
108 }
109 else
110 {
111 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m);
112 }
113
114 if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
115 {
116 const unsigned int src2_dim0 = src2->dimension(0);
117 const unsigned int src2_dim1 = src2->dimension(1);
118
119 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
120 if(gemm_info.broadcast_bias)
121 {
122 ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
123 }
124 else
125 {
126 ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
127 }
128 }
129
130 if(dst->total_size() != 0)
131 {
132 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
133 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
134 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
135 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");
136 }
137
138 return Status{};
139 }
140
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)141 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
142 const GEMMRHSMatrixInfo &rhs_info,
143 const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
144 {
145 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
146 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
147 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
148 bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
149
150 Window win{};
151 Window win_out{};
152 bool window_changed = false;
153
154 // In case both input and dst have to be reinterpreted as 3D tensors,
155 // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
156 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
157 {
158 reinterpret_output_as_3d = false;
159 }
160
161 // dst tensor auto initialization if not yet initialized
162 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
163
164 TensorInfo tmp_info(*dst);
165
166 if(reinterpret_output_as_3d)
167 {
168 // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
169 // the window needs to be constructed on the 2D collapsed version of the tensor
170 TensorShape tmp_shape(dst->tensor_shape());
171 tmp_shape.collapse(2U, 1U);
172 tmp_info.set_tensor_shape(tmp_shape);
173 }
174
175 // Configure kernel window
176 num_elems_processed_per_iteration_x = rhs_info.n0;
177 num_elems_processed_per_iteration_y = lhs_info.m0;
178
179 win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
180 win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
181
182 AccessWindowStatic src0_access(src0, 0, 0,
183 src0->dimension(0),
184 src0->dimension(1));
185 AccessWindowStatic src1_access(src1, 0, 0,
186 ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
187 src1->dimension(1));
188 AccessWindowStatic dst_access(dst, 0, 0,
189 dst->dimension(0),
190 dst->dimension(1));
191
192 if(src2 != nullptr)
193 {
194 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
195
196 AccessWindowStatic src2_access(src2, 0, 0,
197 ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
198 src2->dimension(1));
199
200 window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop
201 update_window_and_padding(win_out, dst_access); // window used to update the padding requirements of dst tensor
202 }
203 else
204 {
205 window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop
206 update_window_and_padding(win_out, dst_access); // window used to update the padding requirements of dst tensor
207 }
208
209 // Collapse along the Z direction
210 // This collapse needs to be here in order to tune the Z dimension of LWS
211 Window collapsed = win;
212 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
213 collapsed = win.collapse(win, dimension_to_collapse);
214
215 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
216 return std::make_pair(err, collapsed);
217 }
218 } // namespace
219
ClGemmMatrixMultiplyNativeKernel()220 ClGemmMatrixMultiplyNativeKernel::ClGemmMatrixMultiplyNativeKernel()
221 {
222 _type = CLKernelType::GEMM;
223 }
224
configure(const CLCompileContext & compile_context,ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * src2,ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)225 void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha,
226 float beta,
227 const GEMMLHSMatrixInfo &lhs_info,
228 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
229 {
230 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
231
232 // dst tensor auto initialization if not yet initialized
233 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
234
235 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
236
237 auto padding_info = get_padding_info({ src0, dst });
238 _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
239 _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
240 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
241 _add_bias = src2 != nullptr;
242 _num_post_op_args = gemm_info.post_ops.total_num_arguments();
243
244 // In case both input and dst have to be reinterpreted as 3D tensors,
245 // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
246 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
247 {
248 _reinterpret_input_as_3d = false;
249 _reinterpret_output_as_3d = false;
250 }
251
252 // Check if we need to slide the matrix B
253 const unsigned int num_dimensions_src0 = src0->num_dimensions();
254 _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0);
255
256 ElementsProcessed num_elements_processed{};
257
258 // Configure kernel window
259 auto win_config = validate_and_configure_window(src0, src1, src2 != nullptr ? src2 : nullptr, dst, lhs_info, rhs_info, gemm_info, num_elements_processed);
260 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
261 IClKernel::configure_internal(win_config.second);
262
263 // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
264 // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
265 // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m
266 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
267
268 const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1);
269 const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2);
270
271 // 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.
272 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
273 const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
274
275 // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
276 // NOTE: This might have implications on heuristics and performance
277 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
278 _m = internal_m;
279 _n = gemm_info.n;
280 _k = gemm_info.k;
281
282 // Create build options
283 CLBuildOptions build_opts;
284 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
285 build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
286 build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
287 build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
288 build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
289 build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
290 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
291 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
292 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
293 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
294 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
295 build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
296 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
297 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
298 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
299 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
300 // If post_ops are used, then we disable the use of gemm_info.activation_info
301 if(gemm_info.post_ops.size() > 0)
302 {
303 post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops);
304 }
305 else
306 {
307 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
308 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
309 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
310 }
311
312 std::string kernel_name("gemm_mm_native");
313 post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops);
314
315 // A macro guard to compile ONLY the kernel of interest
316 build_opts.add_option("-D" + upper_string(kernel_name));
317
318 // Create kernel
319 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
320
321 // Set config_id for enabling LWS tuning
322 _config_id = kernel_name;
323 _config_id += "_";
324 _config_id += (_add_bias ? "add_bias_" : "");
325 _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
326 _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
327 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
328 _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
329 _config_id += lower_string(string_from_data_type(src0->data_type()));
330 _config_id += "_";
331 _config_id += support::cpp11::to_string(dst->dimension(1));
332 _config_id += "_";
333 _config_id += support::cpp11::to_string(dst->dimension(0));
334 _config_id += "_";
335 _config_id += support::cpp11::to_string(gemm_info.k);
336 _config_id += "_";
337 _config_id += support::cpp11::to_string(dst->dimension(2));
338 _config_id += "_";
339 _config_id += support::cpp11::to_string(lhs_info.m0);
340 _config_id += "_";
341 _config_id += support::cpp11::to_string(rhs_info.n0);
342 _config_id += "_";
343 _config_id += support::cpp11::to_string(rhs_info.k0);
344
345 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
346 }
347
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)348 Status ClGemmMatrixMultiplyNativeKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
349 const GEMMLHSMatrixInfo &lhs_info,
350 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
351 {
352 ElementsProcessed num_elements_processed{};
353 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
354 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
355 src1->clone().get(),
356 src2 != nullptr ? src2->clone().get() : nullptr,
357 dst->clone().get(),
358 lhs_info,
359 rhs_info,
360 gemm_info,
361 num_elements_processed)
362 .first);
363
364 return Status{};
365 }
366
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)367 void ClGemmMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
368 {
369 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
370 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
371
372 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
373 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
374 const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
375 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
376
377 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
378 ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
379
380 if(src1->info()->num_dimensions() < 3)
381 {
382 // The stride_z for matrix B must be zero if we do not slice
383 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
384 }
385
386 Window slice = window.first_slice_window_3D();
387 Window slice_matrix_b = slice;
388
389 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
390 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
391
392 if(_reinterpret_input_as_3d)
393 {
394 // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
395 unsigned int idx0;
396 if(_add_bias)
397 {
398 idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + (7 + _num_post_op_args);
399 }
400 else
401 {
402 idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + (6 + _num_post_op_args);
403 }
404 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
405 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
406 }
407
408 if(_reinterpret_output_as_3d)
409 {
410 // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor
411 unsigned int idx0;
412 if(_add_bias)
413 {
414 idx0 = (4 + _num_post_op_args) * num_arguments_per_2D_tensor() + 7 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args;
415 }
416 else
417 {
418 idx0 = (3 + _num_post_op_args) * num_arguments_per_2D_tensor() + 6 + (_reinterpret_input_as_3d ? 1 : 0) + _num_post_op_args;
419 }
420 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
421 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
422 }
423
424 do
425 {
426 Window slice_b = slice;
427 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
428 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
429 if(!_slide_matrix_b)
430 {
431 slice_b = slice_matrix_b;
432 }
433
434 unsigned int idx = 0;
435 add_2D_tensor_argument(idx, src0, slice);
436 add_2D_tensor_argument(idx, src1, slice_b);
437 if(_add_bias)
438 {
439 add_2D_tensor_argument(idx, src2, slice);
440 }
441 add_2D_tensor_argument(idx, dst, slice);
442 // post op argument buffers
443 for(size_t i = 0; i < _num_post_op_args; ++i)
444 {
445 const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
446 add_2D_tensor_argument(idx, post_op_arg, slice);
447 }
448 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
449 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
450 if(_add_bias)
451 {
452 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
453 }
454 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
455 // post op argument stride_z
456 for(size_t i = 0; i < _num_post_op_args; ++i)
457 {
458 const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
459 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
460 }
461
462 // Pass m, n and k at runtime
463 _kernel.setArg<cl_int>(idx++, _m);
464 _kernel.setArg<cl_int>(idx++, _n);
465 _kernel.setArg<cl_int>(idx++, _k);
466
467 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
468 }
469 while(window.slide_window_slice_3D(slice));
470 }
471 } // namespace kernels
472 } // namespace opencl
473 } // namespace arm_compute
474