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/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
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38
39 #include "support/Cast.h"
40 #include "support/StringSupport.h"
41
42 namespace arm_compute
43 {
44 namespace opencl
45 {
46 namespace kernels
47 {
48 using namespace misc::shape_calculator;
49
50 namespace
51 {
52 using ElementsProcessed = Steps;
53
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)54 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst,
55 const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
58 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
60 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
62 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose);
63 ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose);
64 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0);
65 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");
66 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 > 16);
67 ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8);
68 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");
69 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for quantized GEMM");
70
71 const int m = gemm_info.m();
72 const int n = gemm_info.n();
73 const int k = gemm_info.k();
74
75 TensorShape tensor_shape0{ src0->tensor_shape() };
76 tensor_shape0.set(0, k);
77 tensor_shape0.set(1, m);
78
79 TensorShape tensor_shape1{ src1->tensor_shape() };
80 tensor_shape1.set(0, n);
81 tensor_shape1.set(1, k);
82
83 const TensorInfo tensor_info0 = src0->clone()->set_tensor_shape(tensor_shape0);
84 const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
85
86 const TensorInfo tensor_info_reshaped0 = src0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
87 const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
88
89 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src0, &tensor_info_reshaped0);
90 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
91
92 if(dst->total_size() != 0)
93 {
94 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(compute_mm_shape(*src0, *src1, gemm_info));
95 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
96 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32);
97 }
98
99 return Status{};
100 }
101
validate_and_configure_window(const ITensorInfo * src0,const ITensorInfo * src1,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info,ElementsProcessed & num_elements_processed)102 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst,
103 const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info,
104 ElementsProcessed &num_elements_processed)
105 {
106 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
107 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
108 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
109
110 // dst tensor auto initialization if not yet initialized
111 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(compute_mm_shape(*src0, *src1, gemm_info)).set_data_type(DataType::S32));
112
113 TensorInfo tmp_info(*dst);
114 if(reinterpret_output_as_3d)
115 {
116 // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
117 // the window needs to be constructed on the 2D collapsed version of the tensor
118 TensorShape tmp_shape(dst->tensor_shape());
119 tmp_shape.collapse(2U, 1U);
120 tmp_info.set_tensor_shape(tmp_shape);
121 }
122
123 // Configure kernel window
124 num_elems_processed_per_iteration_x = rhs_info.n0;
125 num_elems_processed_per_iteration_y = lhs_info.m0;
126 Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
127
128 // Collapse along the Z direction
129 // This collapse needs to be here in order to tune the Z dimension of LWS
130 Window collapsed = win;
131 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
132 collapsed = win.collapse(win, dimension_to_collapse);
133
134 return std::make_pair(Status{}, collapsed);
135 }
136 } // namespace
137
ClGemmLowpMatrixMultiplyReshapedKernel()138 ClGemmLowpMatrixMultiplyReshapedKernel::ClGemmLowpMatrixMultiplyReshapedKernel()
139 {
140 _type = CLKernelType::GEMM;
141 }
142
configure(const CLCompileContext & compile_context,const ITensorInfo * src0,const ITensorInfo * src1,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)143 void ClGemmLowpMatrixMultiplyReshapedKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst,
144 const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
145 {
146 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
147 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst, lhs_info, rhs_info, gemm_info));
148
149 _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
150 _k = gemm_info.k();
151 _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
152
153 // Check if we need to slide the matrix B
154 const unsigned int num_dimensionssrc0 = src0->num_dimensions();
155 _slide_matrix_b = (src1->num_dimensions() >= num_dimensionssrc0);
156
157 auto padding_info = get_padding_info({ src0, src1, dst });
158 ElementsProcessed num_elements_processed{};
159
160 // Configure kernel window
161 auto win_config = validate_and_configure_window(src0, src1, dst, lhs_info, rhs_info, gemm_info, num_elements_processed);
162 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
163 ICLKernel::configure_internal(win_config.second);
164
165 // 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.
166 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m() : dst->dimension(1);
167
168 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
169 const unsigned int partial_store_n0 = gemm_info.n() % rhs_info.n0;
170
171 // Create build options
172 CLBuildOptions build_opts;
173 build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
174 build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(dst->dimension(1)));
175 build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(dst->dimension(2)));
176 build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
177 build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
178 build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
179 build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
180 build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m()));
181 build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
182 build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
183 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
184 build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
185 build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
186 build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
187 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
188 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(src0->data_type()));
189 build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
190 build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
191
192 std::string kernel_name("gemmlowp_mm_reshaped_");
193 kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
194 kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
195
196 // A macro guard to compile ONLY the kernel of interest
197 build_opts.add_option("-D" + upper_string(kernel_name));
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(dst->dimension(1));
209 _config_id += "_";
210 _config_id += support::cpp11::to_string(dst->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(dst->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 * src0,const ITensorInfo * src1,const ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMReshapeInfo & gemm_info)233 Status ClGemmLowpMatrixMultiplyReshapedKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, 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(src0, src1, dst, lhs_info, rhs_info, gemm_info));
238 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
239 src1->clone().get(),
240 dst->clone().get(),
241 lhs_info,
242 rhs_info,
243 gemm_info,
244 num_elements_processed)
245 .first);
246
247 return Status{};
248 }
249
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)250 void ClGemmLowpMatrixMultiplyReshapedKernel::run_op(ITensorPack &tensors, 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 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
256 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
257 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
258
259 if(src1->info()->num_dimensions() < 3)
260 {
261 // The stride_z for matrix B must be zero if we do not slice
262 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
263 }
264
265 Window slice = window.first_slice_window_3D();
266 Window slice_matrix_b = slice;
267
268 slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
269 slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
270
271 if(_reinterpret_output_as_3d)
272 {
273 // Pass bottom paddings to the kernel if the dst has to be reinterpreted as 3D tensor
274 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 4;
275 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
276 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
277 }
278
279 do
280 {
281 Window slice_b = slice;
282 // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
283 // This scenario can happen when the matrix multiplication is used to perform a convolution operation
284 if(!_slide_matrix_b)
285 {
286 slice_b = slice_matrix_b;
287 }
288
289 unsigned int idx = 0;
290 add_2D_tensor_argument(idx, src0, slice);
291 add_2D_tensor_argument(idx, src1, slice_b);
292 add_2D_tensor_argument(idx, dst, slice);
293 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
294 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
295 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
296 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
297 enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
298 }
299 while(window.slide_window_slice_3D(slice));
300 }
301 } // namespace kernels
302 } // namespace opencl
303 } // namespace arm_compute
304