blob: 52c8cd4fd51377239a74c3eae62a2c65e24f10cd [file] [log] [blame]
Georgios Pinitas856f66e2021-04-22 21:13:21 +01001/*
2 * Copyright (c) 2018-2021 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 */
Georgios Pinitas7891a732021-08-20 21:39:25 +010024#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h"
Georgios Pinitas856f66e2021-04-22 21:13:21 +010025
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"
Georgios Pinitas856f66e2021-04-22 21:13:21 +010035#include "src/core/CL/CLUtils.h"
36#include "src/core/CL/CLValidate.h"
SiCongLi1af54162021-10-06 15:25:57 +010037#include "src/core/experimental/PostOp.h"
Georgios Pinitas856f66e2021-04-22 21:13:21 +010038#include "src/core/helpers/AutoConfiguration.h"
39#include "src/core/helpers/WindowHelpers.h"
40#include "src/core/utils/helpers/float_ops.h"
Georgios Pinitas7891a732021-08-20 21:39:25 +010041#include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
Georgios Pinitas856f66e2021-04-22 21:13:21 +010042#include "support/Cast.h"
43#include "support/StringSupport.h"
44
Georgios Pinitas856f66e2021-04-22 21:13:21 +010045namespace arm_compute
46{
47namespace opencl
48{
49namespace kernels
50{
51namespace
52{
53using ElementsProcessed = Steps;
54
SiCongLi1af54162021-10-06 15:25:57 +010055const auto post_op_utils = experimental::PostOpCLKernelUtils(
56{
57 // PostOp sequence -> {Kernel Postfix, PostOp Slots}
58 { {}, { "", {} } },
59 { { experimental::PostOpType::Activation }, { "", { 1 } } },
ramelg016049eda2021-10-29 10:52:53 +010060
SiCongLi1af54162021-10-06 15:25:57 +010061 { { experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 2 } } },
ramelg016049eda2021-10-29 10:52:53 +010062 { { experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 2 } } },
63
SiCongLi1af54162021-10-06 15:25:57 +010064 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_Add }, { "_post_act_eltwise_op_act", { 1, 2 } } },
ramelg016049eda2021-10-29 10:52:53 +010065 { { experimental::PostOpType::Activation, experimental::PostOpType::Eltwise_PRelu }, { "_post_act_eltwise_op_act", { 1, 2 } } },
66
SiCongLi1af54162021-10-06 15:25:57 +010067 { { experimental::PostOpType::Eltwise_Add, experimental::PostOpType::Activation }, { "_post_act_eltwise_op_act", { 2, 3 } } },
ramelg016049eda2021-10-29 10:52:53 +010068 { { 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 } } }
SiCongLi1af54162021-10-06 15:25:57 +010072});
ramelg016049eda2021-10-29 10:52:53 +010073
Georgios Pinitas856f66e2021-04-22 21:13:21 +010074Status 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));
SiCongLi1af54162021-10-06 15:25:57 +010097 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");
Georgios Pinitas856f66e2021-04-22 21:13:21 +010098
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);
SiCongLi1af54162021-10-06 15:25:57 +0100141 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");
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100142 }
143
144 return Status{};
145}
146
147std::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{
Giorgio Arenabde2f352021-09-07 14:15:28 +0100151 ARM_COMPUTE_UNUSED(src0, src1, src2);
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100152 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
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100156 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
Giorgio Arenabde2f352021-09-07 14:15:28 +0100171 Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100172
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
Giorgio Arenabde2f352021-09-07 14:15:28 +0100179 return std::make_pair(Status{}, collapsed);
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100180}
181} // namespace
182
Giorgio Arena4a95bba2021-06-28 11:00:27 +0100183ClGemmMatrixMultiplyReshapedKernel::ClGemmMatrixMultiplyReshapedKernel()
184{
185 _type = CLKernelType::GEMM;
186}
187
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100188void ClGemmMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context,
Giorgio Arenabde2f352021-09-07 14:15:28 +0100189 const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta,
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100190 const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
191{
192 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
193
Giorgio Arenabde2f352021-09-07 14:15:28 +0100194 // 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
SiCongLieb8bd812021-10-29 15:05:49 +0100197 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
198
Giorgio Arenabde2f352021-09-07 14:15:28 +0100199 auto padding_info = get_padding_info({ src0, src1, src2, dst });
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100200 _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 _k = gemm_info.k;
SiCongLi1af54162021-10-06 15:25:57 +0100205 _num_post_op_args = gemm_info.post_ops.total_num_arguments();
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100206
207 // Check if we need to slide the matrix B
208 const unsigned int num_dimensions_src0 = src0->num_dimensions();
209 _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0);
210
211 ElementsProcessed num_elements_processed{};
212
213 // Configure kernel window
Giorgio Arenabde2f352021-09-07 14:15:28 +0100214 auto win_config = validate_and_configure_window(src0->clone().get(),
215 src1->clone().get(),
216 (src2 != nullptr) ? src2->clone().get() : nullptr,
217 dst->clone().get(),
218 lhs_info,
219 rhs_info,
220 gemm_info,
221 num_elements_processed);
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100222 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
223 ICLKernel::configure_internal(win_config.second);
224
225 const bool enable_mixed_precision = gemm_info.fp_mixed_precision;
226 const DataType data_type = src0->data_type();
227
228 // 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.
229 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
230
231 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
232 const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
233
234 // Create build options
235 CLBuildOptions build_opts;
236 build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
237 build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
238 build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
239 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
240 build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
241 build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
242 build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
243 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
244 build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
245 build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
246 build_opts.add_option_if(lhs_info.transpose, "-DLHS_TRANSPOSE");
247 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100248 build_opts.add_option_if(enable_mixed_precision, "-DMIXED_PRECISION");
249 build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
250 build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(src1->dimension(1)));
251 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
252 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)));
253 build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m));
254 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
255 build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
256 build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
257 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
258 build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
259 build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
260 build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
261 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
262 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
SiCongLi1af54162021-10-06 15:25:57 +0100263 // If post_ops are used, then we disable the use of gemm_info.activation_info
264 if(gemm_info.post_ops.size() > 0)
265 {
266 post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.post_ops);
267 }
268 else
269 {
270 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
271 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
272 build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
273 }
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100274
275 std::string kernel_name("gemm_mm_reshaped_");
276 kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
277 kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
278 kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
SiCongLi1af54162021-10-06 15:25:57 +0100279 post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops);
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100280
281 // Create kernel
282 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
283
284 // Set config_id for enabling LWS tuning
285 _config_id = kernel_name;
286 _config_id += "_";
287 _config_id += (_add_bias ? "add_bias_" : "");
288 _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
289 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
290 _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
291 _config_id += lower_string(string_from_data_type(src0->data_type()));
292 _config_id += "_";
293 _config_id += (enable_mixed_precision ? "mixed_precision_" : "");
294 _config_id += support::cpp11::to_string(dst->dimension(1));
295 _config_id += "_";
296 _config_id += support::cpp11::to_string(dst->dimension(0));
297 _config_id += "_";
298 _config_id += support::cpp11::to_string(gemm_info.k);
299 _config_id += "_";
300 _config_id += support::cpp11::to_string(dst->dimension(2));
301 _config_id += "_";
302 _config_id += support::cpp11::to_string(lhs_info.m0);
303 _config_id += "_";
304 _config_id += support::cpp11::to_string(rhs_info.n0);
305 _config_id += "_";
306 _config_id += support::cpp11::to_string(lhs_info.k0);
307 _config_id += "_";
308 _config_id += support::cpp11::to_string(lhs_info.v0);
309 _config_id += "_";
310 _config_id += support::cpp11::to_string(rhs_info.h0);
311 _config_id += "_";
312 _config_id += support::cpp11::to_string(lhs_info.interleave);
313 _config_id += "_";
314 _config_id += support::cpp11::to_string(rhs_info.interleave);
315
316 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
317}
318
319Status ClGemmMatrixMultiplyReshapedKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
320 const GEMMLHSMatrixInfo &lhs_info,
321 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
322{
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100323 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100324 return Status{};
325}
326
327void ClGemmMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
328{
329 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
330 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
331
332 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
333 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
334 const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
335 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
336
337 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
338 ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
339
340 if(src1->info()->num_dimensions() < 3)
341 {
342 // The stride_z for matrix B must be zero if we do not slice
343 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
344 }
345
346 Window slice = window.first_slice_window_3D();
347 Window slice_matrix_b = slice;
348
349 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
350 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
351
352 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
353
354 cl::Image2D src1_image2d;
355
356 if(_export_to_cl_image)
357 {
358 const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2));
359 const size_t image_row_pitch = src1->info()->strides_in_bytes()[1];
360
361 src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch);
362 }
363
364 do
365 {
366 Window slice_b = slice;
367 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
368 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
369 if(!_slide_matrix_b)
370 {
371 slice_b = slice_matrix_b;
372 }
373
374 unsigned int idx = 0;
375
376 // LHS buffer
377 add_2D_tensor_argument(idx, src0, slice);
378
379 // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
380 if(_export_to_cl_image)
381 {
382 _kernel.setArg(idx++, src1_image2d);
383 }
384 else
385 {
386 add_2D_tensor_argument(idx, src1, slice_b);
387 }
388
389 // Bias buffer (_add_bias == true)
390 add_2D_tensor_argument_if(_add_bias, idx, src2, slice);
391
392 // dst buffer
393 add_2D_tensor_argument(idx, dst, slice);
394
SiCongLi1af54162021-10-06 15:25:57 +0100395 // post op argument buffers
396 for(size_t i = 0; i < _num_post_op_args; ++i)
397 {
398 const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
399 add_2D_tensor_argument(idx, post_op_arg, slice);
400 }
401
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100402 // K dimension (not used if _export_to_cl_image == true)
403 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
404
405 // LHS stride_z
406 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
407
408 // RHS stride_z (not used if _export_to_cl_image == true)
409 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
410
411 // Bias stride_z (if _add_bias == true)
412 if(_add_bias)
413 {
414 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
415 }
416
417 // dst stride_z
418 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
419
SiCongLi1af54162021-10-06 15:25:57 +0100420 // post op argument stride_z
421 for(size_t i = 0; i < _num_post_op_args; ++i)
422 {
423 const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
424 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
425 }
Georgios Pinitas856f66e2021-04-22 21:13:21 +0100426 // Cross-plan padding (if _reinterpret_output_as_3d = true)
427 if(_reinterpret_output_as_3d)
428 {
429 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad));
430 }
431
432 // Dispatch kernel
433 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
434 }
435 while(window.slide_window_slice_3D(slice));
436}
437} // namespace kernels
438} // namespace opencl
439} // namespace arm_compute