1 /*
2 * Copyright (c) 2017-2020 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/CLGEMMLowpReductionKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/ICLTensor.h"
28 #include "arm_compute/core/KernelDescriptors.h"
29 #include "src/core/AccessWindowStatic.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 #include "support/StringSupport.h"
33
34 namespace arm_compute
35 {
36 namespace
37 {
validate_arguments_matrix_a_reduction(const ITensorInfo * input,const ITensorInfo * output)38 Status validate_arguments_matrix_a_reduction(const ITensorInfo *input, const ITensorInfo *output)
39 {
40 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
41 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8);
42
43 if(output->total_size() > 0)
44 {
45 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
46 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(0) != input->dimension(1), "Output vector must have length equal to the number of rows of the input matrix");
47 }
48 return Status{};
49 }
50
validate_arguments_matrix_b_reduction(const ITensorInfo * input,const ITensorInfo * output)51 Status validate_arguments_matrix_b_reduction(const ITensorInfo *input, const ITensorInfo *output)
52 {
53 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
54 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL);
55
56 if(output->total_size() > 0)
57 {
58 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
59 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(0) != input->dimension(0), "Output vector must have length equal to the number of columns of the input matrix");
60 }
61 return Status{};
62 }
63 } // namespace
64
ICLGEMMLowpReductionKernel()65 ICLGEMMLowpReductionKernel::ICLGEMMLowpReductionKernel()
66 : _input(), _output()
67 {
68 }
69
configure(const ICLTensor * mtx_a,ICLTensor * vector_sum_row,const GEMMLowpReductionKernelInfo & info)70 void CLGEMMLowpMatrixAReductionKernel::configure(const ICLTensor *mtx_a, ICLTensor *vector_sum_row, const GEMMLowpReductionKernelInfo &info)
71 {
72 configure(CLKernelLibrary::get().get_compile_context(), mtx_a, vector_sum_row, info);
73 }
74
configure(const CLCompileContext & compile_context,const ICLTensor * mtx_a,ICLTensor * vector_sum_row,const GEMMLowpReductionKernelInfo & info)75 void CLGEMMLowpMatrixAReductionKernel::configure(const CLCompileContext &compile_context, const ICLTensor *mtx_a, ICLTensor *vector_sum_row, const GEMMLowpReductionKernelInfo &info)
76 {
77 // Perform validate step
78 ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_a, vector_sum_row);
79 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_a_reduction(mtx_a->info(), vector_sum_row->info()));
80
81 // Output auto initialization if not yet initialized
82 auto_init_if_empty(*vector_sum_row->info(), TensorShape(mtx_a->info()->dimension(1)), 1, DataType::S32);
83
84 auto padding_info = get_padding_info({ mtx_a, vector_sum_row });
85
86 _input = mtx_a;
87 _output = vector_sum_row;
88
89 // Set the arguments to pass at compile time
90 CLBuildOptions build_opts;
91 build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(mtx_a->info()->dimension(0)));
92 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(mtx_a->info()->data_type()));
93 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(mtx_a->info()->data_type()));
94 build_opts.add_option_if(info.mul_by_scalar, "-DSCALAR=" + support::cpp11::to_string(info.scalar));
95
96 const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device());
97
98 std::string kernel_name = "gemmlowp_matrix_a_reduction" + std::string(is_dot8_supported ? "_dot8" : "");
99
100 // Create kernel
101 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
102
103 // Configure kernel window
104 // This kernel does not need padding
105 Window win = calculate_max_window(*vector_sum_row->info(), Steps());
106 ICLKernel::configure_internal(win);
107
108 _config_id = kernel_name;
109 _config_id += "_";
110 _config_id += support::cpp11::to_string(_input->info()->dimension(0));
111 _config_id += "_";
112 _config_id += support::cpp11::to_string(_input->info()->dimension(1));
113 _config_id += "_";
114 _config_id += support::cpp11::to_string(_input->info()->dimension(2));
115
116 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
117 }
118
validate(const ITensorInfo * mtx_a,const ITensorInfo * vector_sum_row,const GEMMLowpReductionKernelInfo & info)119 Status CLGEMMLowpMatrixAReductionKernel::validate(const ITensorInfo *mtx_a, const ITensorInfo *vector_sum_row, const GEMMLowpReductionKernelInfo &info)
120 {
121 ARM_COMPUTE_UNUSED(info);
122 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_a_reduction(mtx_a, vector_sum_row));
123
124 return Status{};
125 }
126
run(const Window & window,cl::CommandQueue & queue)127 void CLGEMMLowpMatrixAReductionKernel::run(const Window &window, cl::CommandQueue &queue)
128 {
129 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
130 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
131
132 Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimY);
133 Window slice_in = collapsed.first_slice_window_2D();
134 Window slice_out = collapsed.first_slice_window_2D();
135
136 // Setup input slice. Its dimensions are increased in the cl kernel.
137 slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
138 slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
139 slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
140
141 do
142 {
143 unsigned int idx = 0;
144 add_3D_tensor_argument(idx, _input, slice_in);
145 add_2D_tensor_argument(idx, _output, slice_out);
146 enqueue(queue, *this, slice_out, lws_hint());
147 }
148 while(collapsed.slide_window_slice_2D(slice_out));
149 }
150
configure(const ICLTensor * mtx_b,ICLTensor * vector_sum_col,const GEMMLowpReductionKernelInfo & info)151 void CLGEMMLowpMatrixBReductionKernel::configure(const ICLTensor *mtx_b, ICLTensor *vector_sum_col, const GEMMLowpReductionKernelInfo &info)
152 {
153 configure(CLKernelLibrary::get().get_compile_context(), mtx_b, vector_sum_col, info);
154 }
155
configure(const CLCompileContext & compile_context,const ICLTensor * mtx_b,ICLTensor * vector_sum_col,const GEMMLowpReductionKernelInfo & info)156 void CLGEMMLowpMatrixBReductionKernel::configure(const CLCompileContext &compile_context, const ICLTensor *mtx_b, ICLTensor *vector_sum_col, const GEMMLowpReductionKernelInfo &info)
157 {
158 ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_b, vector_sum_col);
159 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_b_reduction(mtx_b->info(), vector_sum_col->info()));
160
161 _input = mtx_b;
162 _output = vector_sum_col;
163
164 // Output auto initialization if not yet initialized
165 auto_init_if_empty(*_output->info(), TensorShape(mtx_b->info()->dimension(0)), 1, DataType::S32);
166
167 auto padding_info = get_padding_info({ mtx_b, vector_sum_col });
168
169 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, mtx_b->info()->dimension(0));
170
171 // Set the arguments to pass at compile time
172 CLBuildOptions build_opts;
173 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
174 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mtx_b->info()->dimension(0) % num_elems_processed_per_iteration));
175 build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(0)));
176 build_opts.add_option("-DROWS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(1)));
177 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(mtx_b->info()->data_type()));
178 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(mtx_b->info()->data_type()));
179 build_opts.add_option_if(info.mul_by_scalar, "-DSCALAR=" + support::cpp11::to_string(info.scalar));
180
181 // Create kernel
182 _kernel = create_kernel(compile_context, "gemmlowp_matrix_b_reduction", build_opts.options());
183
184 // Configure kernel window
185 Window win = calculate_max_window(*_output->info(), Steps(num_elems_processed_per_iteration));
186 ICLKernel::configure_internal(win);
187
188 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
189 }
190
validate(const ITensorInfo * mtx_b,const ITensorInfo * vector_sum_col,const GEMMLowpReductionKernelInfo & info)191 Status CLGEMMLowpMatrixBReductionKernel::validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, const GEMMLowpReductionKernelInfo &info)
192 {
193 ARM_COMPUTE_UNUSED(info);
194 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_b_reduction(mtx_b, vector_sum_col));
195
196 return Status{};
197 }
198
run(const Window & window,cl::CommandQueue & queue)199 void CLGEMMLowpMatrixBReductionKernel::run(const Window &window, cl::CommandQueue &queue)
200 {
201 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
202 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
203
204 Window collapsed = window.collapse_if_possible(IKernel::window(), Window::DimY);
205
206 Window slice_out = collapsed.first_slice_window_2D();
207 Window slice_in = slice_out;
208
209 slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
210 slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
211
212 do
213 {
214 unsigned int idx = 0;
215 add_3D_tensor_argument(idx, _input, slice_in);
216 add_2D_tensor_argument(idx, _output, slice_out);
217 enqueue(queue, *this, slice_out, lws_hint());
218 }
219 while(collapsed.slide_window_slice_2D(slice_out));
220 }
221 } // namespace arm_compute
222