1 /*
2 * Copyright (c) 2019-2022 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/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
36 #include "src/core/AccessWindowStatic.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39
40 #include "support/Cast.h"
41 #include "support/StringSupport.h"
42
43 namespace arm_compute
44 {
45 namespace opencl
46 {
47 namespace kernels
48 {
49 namespace
50 {
51 using ElementsProcessed = Steps;
52
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)53 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
54 const GEMMReshapeInfo &gemm_info)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
57 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
58 if(src0->data_type() == DataType::QASYMM8)
59 {
60 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
61 }
62 else
63 {
64 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL);
65 }
66 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
67 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
68 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
69 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");
70 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
71 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
72 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");
73 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
74
75 const int m = gemm_info.m();
76 const int n = gemm_info.n();
77 const int k = gemm_info.k();
78
79 ARM_COMPUTE_UNUSED(m);
80 ARM_COMPUTE_UNUSED(n);
81 ARM_COMPUTE_UNUSED(k);
82
83 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != static_cast<unsigned int>(k));
84 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) != static_cast<unsigned int>(n));
85 ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(1) != static_cast<unsigned int>(k));
86 if(gemm_info.reinterpret_input_as_3d())
87 {
88 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != static_cast<unsigned int>(m));
89 }
90 else
91 {
92 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != static_cast<unsigned int>(m));
93 }
94
95 if(dst->total_size() != 0)
96 {
97 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
98 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
99 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32);
100 }
101
102 return Status{};
103 }
104
validate_and_configure_window(const ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info,ElementsProcessed & num_elements_processed)105 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
106 const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
107 {
108 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
109 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
110 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
111 bool reinterpret_dst_as_3d = (gemm_info.depth_output_gemm3d() != 0);
112
113 Window win{};
114 bool window_changed = false;
115
116 // In case both input and dst have to be reinterpreted as 3D tensors,
117 // force reinterpret_dst_as_3d to be false.
118 if(reinterpret_input_as_3d == reinterpret_dst_as_3d)
119 {
120 reinterpret_dst_as_3d = false;
121 }
122
123 // dst tensor auto initialization if not yet initialized
124 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)).set_data_type(DataType::S32));
125
126 TensorInfo tmp_info(*dst);
127
128 if(reinterpret_dst_as_3d)
129 {
130 // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
131 // the window needs to be constructed on the 2D collapsed version of the tensor
132 TensorShape tmp_shape(dst->tensor_shape());
133 tmp_shape.collapse(2U, 1U);
134 tmp_info.set_tensor_shape(tmp_shape);
135 }
136
137 // Configure kernel window
138 num_elems_processed_per_iteration_x = rhs_info.n0;
139 num_elems_processed_per_iteration_y = lhs_info.m0;
140
141 win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
142
143 // RHS matrix still needs padding on the X
144 AccessWindowStatic src1_access(src1, 0, 0,
145 ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
146 src1->dimension(1));
147
148 window_changed = update_window_and_padding(win, src1_access); // window used by the execute_window_loop
149
150 // Collapse along the Z direction
151 // This collapse needs to be here in order to tune the Z dimension of LWS
152 Window collapsed = win;
153 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
154 collapsed = win.collapse(win, dimension_to_collapse);
155
156 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
157 return std::make_pair(err, collapsed);
158 }
159 } // namespace
160
ClGemmLowpMatrixMultiplyNativeKernel()161 ClGemmLowpMatrixMultiplyNativeKernel::ClGemmLowpMatrixMultiplyNativeKernel()
162 {
163 _type = CLKernelType::GEMM;
164 }
165
configure(const CLCompileContext & compile_context,const ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)166 void ClGemmLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *dst,
167 const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
168 {
169 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
170
171 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst, lhs_info, rhs_info, gemm_info));
172
173 _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
174 _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
175 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
176
177 // We still need padding on the X dimension for the RHS matrix
178 auto padding_info = get_padding_info({ src0, dst });
179
180 // In case both input and dst have to be reinterpreted as 3D tensors,
181 // force reinterpret_input_as_3d and reinterpret_dst_as_3d to be false.
182 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
183 {
184 _reinterpret_input_as_3d = false;
185 _reinterpret_output_as_3d = false;
186 }
187
188 // Check if we need to slide the matrix B
189 const unsigned int num_dimensions_src0 = src0->num_dimensions();
190 _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0);
191
192 ElementsProcessed num_elements_processed{};
193
194 // Configure kernel window
195 auto win_config = validate_and_configure_window(src0, src1, dst, lhs_info, rhs_info, gemm_info, num_elements_processed);
196 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
197 ICLKernel::configure_internal(win_config.second);
198
199 // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
200 // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
201 // This means that the actual m used by the kernel is given by dst->info()->dimension(1) and not by gemm_info.m
202 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m() : dst->dimension(1);
203 // 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.
204 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
205 const unsigned int partial_store_n0 = gemm_info.n() % rhs_info.n0;
206
207 // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
208 // NOTE: This might have implications on heuristics and performance
209 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
210
211 // Create build options
212 CLBuildOptions build_opts;
213 build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
214 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
215 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
216 build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
217 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
218 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
219 build_opts.add_option("-DM=" + support::cpp11::to_string(src0->dimension(1)));
220 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
221 build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k()));
222 build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
223 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
224 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
225 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
226 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(src0->data_type()));
227 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
228 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
229 std::string kernel_name("gemmlowp_mm_native");
230
231 // A macro guard to compile ONLY the kernel of interest
232 build_opts.add_option("-D" + upper_string(kernel_name));
233
234 // Create kernel
235 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
236
237 // Set config_id for enabling LWS tuning
238 _config_id = kernel_name;
239 _config_id += "_";
240 _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : "";
241 _config_id += "_";
242 _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
243 _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
244 _config_id += support::cpp11::to_string(dst->dimension(1));
245 _config_id += "_";
246 _config_id += support::cpp11::to_string(dst->dimension(0));
247 _config_id += "_";
248 _config_id += support::cpp11::to_string(gemm_info.k());
249 _config_id += "_";
250 _config_id += support::cpp11::to_string(dst->dimension(2));
251 _config_id += "_";
252 _config_id += support::cpp11::to_string(lhs_info.m0);
253 _config_id += "_";
254 _config_id += support::cpp11::to_string(rhs_info.n0);
255 _config_id += "_";
256 _config_id += support::cpp11::to_string(lhs_info.k0);
257
258 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
259 }
260
validate(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)261 Status ClGemmLowpMatrixMultiplyNativeKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
262 const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
263 {
264 ElementsProcessed num_elements_processed{};
265 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, dst, lhs_info, rhs_info, gemm_info));
266 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
267 src1->clone().get(),
268 dst->clone().get(),
269 lhs_info,
270 rhs_info,
271 gemm_info,
272 num_elements_processed)
273 .first);
274
275 return Status{};
276 }
277
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)278 void ClGemmLowpMatrixMultiplyNativeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
279 {
280 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
281 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
282
283 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
284 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
285 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
286
287 if(src1->info()->num_dimensions() < 3)
288 {
289 // The stride_z for matrix B must be zero if we do not slice
290 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
291 }
292
293 Window slice = window.first_slice_window_3D();
294 Window slice_matrix_b = slice;
295
296 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
297 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
298
299 if(_reinterpret_input_as_3d)
300 {
301 // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
302 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3;
303 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
304 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
305 }
306
307 if(_reinterpret_output_as_3d)
308 {
309 // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
310 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
311 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
312 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
313 }
314
315 do
316 {
317 Window slice_b = slice;
318 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
319 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
320 if(!_slide_matrix_b)
321 {
322 slice_b = slice_matrix_b;
323 }
324
325 unsigned int idx = 0;
326 add_2D_tensor_argument(idx, src0, slice);
327 add_2D_tensor_argument(idx, src1, slice_b);
328 add_2D_tensor_argument(idx, dst, slice);
329 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
330 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
331 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
332 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
333 }
334 while(window.slide_window_slice_3D(slice));
335 }
336 } // namespace kernels
337 } // namespace opencl
338 } // namespace arm_compute
339