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/CLGEMMLowpMatrixMultiplyNativeKernel.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/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "support/StringSupport.h"
39
40 namespace arm_compute
41 {
42 using namespace misc::shape_calculator;
43
44 namespace
45 {
46 using ElementsProcessed = Steps;
47
validate_arguments(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)48 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
49 const GEMMReshapeInfo &gemm_info)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
52 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
53 if(input0->data_type() == DataType::QASYMM8)
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
56 }
57 else
58 {
59 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
60 }
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
63 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
64 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");
65 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
66 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
67 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");
68 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
69
70 const int m = gemm_info.m();
71 const int n = gemm_info.n();
72 const int k = gemm_info.k();
73
74 ARM_COMPUTE_UNUSED(m);
75 ARM_COMPUTE_UNUSED(n);
76 ARM_COMPUTE_UNUSED(k);
77
78 ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k));
79 ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != static_cast<unsigned int>(n));
80 ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(1) != static_cast<unsigned int>(k));
81 if(gemm_info.reinterpret_input_as_3d())
82 {
83 ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m));
84 }
85 else
86 {
87 ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m));
88 }
89
90 if(output->total_size() != 0)
91 {
92 const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
93 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
94 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
95 }
96
97 return Status{};
98 }
99
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)100 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
101 const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
102 {
103 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
104 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
105 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
106 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
107
108 Window win{};
109 bool window_changed = false;
110
111 // In case both input and output have to be reinterpreted as 3D tensors,
112 // force reinterpret_output_as_3d to be false.
113 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
114 {
115 reinterpret_output_as_3d = false;
116 }
117
118 // Output tensor auto initialization if not yet initialized
119 auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)).set_data_type(DataType::S32));
120
121 TensorInfo tmp_info(*output);
122
123 if(reinterpret_output_as_3d)
124 {
125 // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
126 // the window needs to be constructed on the 2D collapsed version of the tensor
127 TensorShape tmp_shape(output->tensor_shape());
128 tmp_shape.collapse(2U, 1U);
129 tmp_info.set_tensor_shape(tmp_shape);
130 }
131
132 // Configure kernel window
133 num_elems_processed_per_iteration_x = rhs_info.n0;
134 num_elems_processed_per_iteration_y = lhs_info.m0;
135
136 win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
137
138 // RHS matrix still needs padding on the X
139 AccessWindowStatic input1_access(input1, 0, 0,
140 ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
141 input1->dimension(1));
142
143 window_changed = update_window_and_padding(win, input1_access); // window used by the execute_window_loop
144
145 // Collapse along the Z direction
146 // This collapse needs to be here in order to tune the Z dimension of LWS
147 Window collapsed = win;
148 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
149 collapsed = win.collapse(win, dimension_to_collapse);
150
151 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
152 return std::make_pair(err, collapsed);
153 }
154 } // namespace
155
CLGEMMLowpMatrixMultiplyNativeKernel()156 CLGEMMLowpMatrixMultiplyNativeKernel::CLGEMMLowpMatrixMultiplyNativeKernel()
157 : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false)
158 {
159 }
160
configure(const ICLTensor * input0,const ICLTensor * input1,ICLTensor * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)161 void CLGEMMLowpMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
162 const GEMMReshapeInfo &gemm_info)
163 {
164 configure(CLKernelLibrary::get().get_compile_context(), input0, input1, output, lhs_info, rhs_info, gemm_info);
165 }
166
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)167 void CLGEMMLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info,
168 const GEMMRHSMatrixInfo &rhs_info,
169 const GEMMReshapeInfo &gemm_info)
170 {
171 ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
172
173 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info));
174
175 _input0 = input0;
176 _input1 = input1;
177 _output = output;
178 _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
179 _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
180 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
181
182 // We still need padding on the X dimension for the RHS matrix
183 auto padding_info = get_padding_info({ input0, output });
184
185 // In case both input and output have to be reinterpreted as 3D tensors,
186 // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
187 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
188 {
189 _reinterpret_input_as_3d = false;
190 _reinterpret_output_as_3d = false;
191 }
192
193 // Check if we need to slide the matrix B
194 const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
195 _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);
196
197 ElementsProcessed num_elements_processed{};
198
199 // Configure kernel window
200 auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
201 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
202 ICLKernel::configure_internal(win_config.second);
203
204 // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
205 // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
206 // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m
207 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m() : output->info()->dimension(1);
208 // 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.
209 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
210 const unsigned int partial_store_n0 = gemm_info.n() % rhs_info.n0;
211
212 // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
213 // NOTE: This might have implications on heuristics and performance
214 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
215
216 // Create build options
217 CLBuildOptions build_opts;
218 build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
219 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
220 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
221 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
222 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
223 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
224 build_opts.add_option("-DM=" + support::cpp11::to_string(input0->info()->dimension(1)));
225 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
226 build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k()));
227 build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
228 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
229 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
230 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
231 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(input0->info()->data_type()));
232 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
233 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
234 std::string kernel_name("gemmlowp_mm_native");
235
236 // Create kernel
237 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
238
239 // Set config_id for enabling LWS tuning
240 _config_id = kernel_name;
241 _config_id += "_";
242 _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
243 _config_id += "_";
244 _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
245 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
246 _config_id += support::cpp11::to_string(output->info()->dimension(1));
247 _config_id += "_";
248 _config_id += support::cpp11::to_string(output->info()->dimension(0));
249 _config_id += "_";
250 _config_id += support::cpp11::to_string(gemm_info.k());
251 _config_id += "_";
252 _config_id += support::cpp11::to_string(output->info()->dimension(2));
253 _config_id += "_";
254 _config_id += support::cpp11::to_string(lhs_info.m0);
255 _config_id += "_";
256 _config_id += support::cpp11::to_string(rhs_info.n0);
257 _config_id += "_";
258 _config_id += support::cpp11::to_string(lhs_info.k0);
259
260 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
261 }
262
validate(const ITensorInfo * input0,const ITensorInfo * input1,const ITensorInfo * output,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)263 Status CLGEMMLowpMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
264 const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
265 {
266 ElementsProcessed num_elements_processed{};
267 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, lhs_info, rhs_info, gemm_info));
268 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
269 input1->clone().get(),
270 output->clone().get(),
271 lhs_info,
272 rhs_info,
273 gemm_info,
274 num_elements_processed)
275 .first);
276
277 return Status{};
278 }
279
run(const Window & window,cl::CommandQueue & queue)280 void CLGEMMLowpMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueue &queue)
281 {
282 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
283 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
284
285 if(_input1->info()->num_dimensions() < 3)
286 {
287 // The stride_z for matrix B must be zero if we do not slice
288 ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
289 }
290
291 Window slice = window.first_slice_window_3D();
292 Window slice_matrix_b = slice;
293
294 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
295 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
296
297 if(_reinterpret_input_as_3d)
298 {
299 // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
300 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3;
301 const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
302 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
303 }
304
305 if(_reinterpret_output_as_3d)
306 {
307 // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
308 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
309 const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
310 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
311 }
312
313 do
314 {
315 Window slice_b = slice;
316 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
317 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
318 if(!_slide_matrix_b)
319 {
320 slice_b = slice_matrix_b;
321 }
322
323 unsigned int idx = 0;
324 add_2D_tensor_argument(idx, _input0, slice);
325 add_2D_tensor_argument(idx, _input1, slice_b);
326 add_2D_tensor_argument(idx, _output, slice);
327 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
328 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
329 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
330 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
331 }
332 while(window.slide_window_slice_3D(slice));
333 }
334 } // namespace arm_compute
335