1 /*
2 * Copyright (c) 2019-2020 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.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/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 #include "support/StringSupport.h"
38
39 namespace arm_compute
40 {
41 using namespace misc::shape_calculator;
42
43 namespace
44 {
45 using ElementsProcessed = Steps;
46
validate_arguments(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)47 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
48 const GEMMReshapeInfo &gemm_info)
49 {
50 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
52 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
54 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
55 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose);
56 ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose);
57 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
58 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");
59 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
60 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
61 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");
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
63
64 const int m = gemm_info.m();
65 const int n = gemm_info.n();
66 const int k = gemm_info.k();
67
68 TensorShape tensor_shape0{ input0->tensor_shape() };
69 tensor_shape0.set(0, k);
70 tensor_shape0.set(1, m);
71
72 TensorShape tensor_shape1{ input1->tensor_shape() };
73 tensor_shape1.set(0, n);
74 tensor_shape1.set(1, k);
75
76 const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
77 const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
78
79 const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
80 const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
81
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
83 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
84
85 if(output->total_size() != 0)
86 {
87 const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
88 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
89 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
90 }
91
92 return Status{};
93 }
94
validate_and_configure_window(ITensorInfo * input0,ITensorInfo * input1,ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info,ElementsProcessed & num_elements_processed)95 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
96 const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
97 {
98 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
99 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
100 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
101
102 // Output tensor auto initialization if not yet initialized
103 auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)).set_data_type(DataType::S32));
104
105 TensorInfo tmp_info(*output);
106 if(reinterpret_output_as_3d)
107 {
108 // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
109 // the window needs to be constructed on the 2D collapsed version of the tensor
110 TensorShape tmp_shape(output->tensor_shape());
111 tmp_shape.collapse(2U, 1U);
112 tmp_info.set_tensor_shape(tmp_shape);
113 }
114
115 // Configure kernel window
116 num_elems_processed_per_iteration_x = rhs_info.n0;
117 num_elems_processed_per_iteration_y = lhs_info.m0;
118 Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
119
120 // Collapse along the Z direction
121 // This collapse needs to be here in order to tune the Z dimension of LWS
122 Window collapsed = win;
123 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
124 collapsed = win.collapse(win, dimension_to_collapse);
125
126 return std::make_pair(Status{}, collapsed);
127 }
128 } // namespace
129
CLGEMMLowpMatrixMultiplyReshapedKernel()130 CLGEMMLowpMatrixMultiplyReshapedKernel::CLGEMMLowpMatrixMultiplyReshapedKernel()
131 : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false)
132 {
133 }
134
configure(const ICLTensor * input0,const ICLTensor * input1,ICLTensor * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)135 void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
136 const GEMMReshapeInfo &gemm_info)
137 {
138 configure(CLKernelLibrary::get().get_compile_context(), input0, input1, output, lhs_info, rhs_info, gemm_info);
139 }
140
configure(const CLCompileContext & compile_context,const ICLTensor * input0,const ICLTensor * input1,ICLTensor * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)141 void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info,
142 const GEMMRHSMatrixInfo &rhs_info,
143 const GEMMReshapeInfo &gemm_info)
144 {
145 ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
146
147 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info));
148
149 _input0 = input0;
150 _input1 = input1;
151 _output = output;
152 _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
153 _k = gemm_info.k();
154 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
155
156 // Check if we need to slide the matrix B
157 const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
158 _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);
159
160 auto padding_info = get_padding_info({ input0, input1, output });
161 ElementsProcessed num_elements_processed{};
162
163 // Configure kernel window
164 auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
165 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
166 ICLKernel::configure_internal(win_config.second);
167
168 // 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.
169 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m() : output->info()->dimension(1);
170
171 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
172 const unsigned int partial_store_n0 = gemm_info.n() % rhs_info.n0;
173
174 // Create build options
175 CLBuildOptions build_opts;
176 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
177 build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
178 build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
179 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
180 build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
181 build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
182 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
183 build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m()));
184 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
185 build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
186 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
187 build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
188 build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
189 build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
190 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
191 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(input0->info()->data_type()));
192 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
193 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
194
195 std::string kernel_name("gemmlowp_mm_reshaped_");
196 kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
197 kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
198
199 // Create kernel
200 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
201
202 // Set config_id for enabling LWS tuning
203 _config_id = kernel_name;
204 _config_id += "_";
205 _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
206 _config_id += "_";
207 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
208 _config_id += support::cpp11::to_string(output->info()->dimension(1));
209 _config_id += "_";
210 _config_id += support::cpp11::to_string(output->info()->dimension(0));
211 _config_id += "_";
212 _config_id += support::cpp11::to_string(gemm_info.k());
213 _config_id += "_";
214 _config_id += support::cpp11::to_string(output->info()->dimension(2));
215 _config_id += "_";
216 _config_id += support::cpp11::to_string(lhs_info.m0);
217 _config_id += "_";
218 _config_id += support::cpp11::to_string(rhs_info.n0);
219 _config_id += "_";
220 _config_id += support::cpp11::to_string(lhs_info.k0);
221 _config_id += "_";
222 _config_id += support::cpp11::to_string(lhs_info.v0);
223 _config_id += "_";
224 _config_id += support::cpp11::to_string(rhs_info.h0);
225 _config_id += "_";
226 _config_id += support::cpp11::to_string(lhs_info.interleave);
227 _config_id += "_";
228 _config_id += support::cpp11::to_string(rhs_info.interleave);
229
230 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
231 }
232
validate(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)233 Status CLGEMMLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
234 const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
235 {
236 ElementsProcessed num_elements_processed{};
237 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, lhs_info, rhs_info, gemm_info));
238 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
239 input1->clone().get(),
240 output->clone().get(),
241 lhs_info,
242 rhs_info,
243 gemm_info,
244 num_elements_processed)
245 .first);
246
247 return Status{};
248 }
249
run(const Window & window,cl::CommandQueue & queue)250 void CLGEMMLowpMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQueue &queue)
251 {
252 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
253 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
254
255 if(_input1->info()->num_dimensions() < 3)
256 {
257 // The stride_z for matrix B must be zero if we do not slice
258 ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
259 }
260
261 Window slice = window.first_slice_window_3D();
262 Window slice_matrix_b = slice;
263
264 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
265 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
266
267 if(_reinterpret_output_as_3d)
268 {
269 // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
270 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 4;
271 const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
272 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
273 }
274
275 do
276 {
277 Window slice_b = slice;
278 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
279 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
280 if(!_slide_matrix_b)
281 {
282 slice_b = slice_matrix_b;
283 }
284
285 unsigned int idx = 0;
286 add_2D_tensor_argument(idx, _input0, slice);
287 add_2D_tensor_argument(idx, _input1, slice_b);
288 add_2D_tensor_argument(idx, _output, slice);
289 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
290 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
291 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
292 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
293 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
294 }
295 while(window.slide_window_slice_3D(slice));
296 }
297 } // namespace arm_compute
298