1 /*
2 * Copyright (c) 2017-2022 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/ClGemmLowpOffsetContributionKernel.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Utils.h"
30 #include "arm_compute/core/Validate.h"
31
32 #include "src/core/helpers/WindowHelpers.h"
33
34 #include "support/Cast.h"
35 #include "support/StringSupport.h"
36
37 namespace arm_compute
38 {
39 namespace opencl
40 {
41 namespace kernels
42 {
43 namespace
44 {
validate_arguments(const ITensorInfo * mm_result,const ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,const ITensorInfo * bias,int32_t a_offset,int32_t b_offset)45 Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
46 int32_t a_offset, int32_t b_offset)
47 {
48 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32);
49
50 if(bias != nullptr)
51 {
52 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
53 ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
54 ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0));
55 }
56
57 // If a_offset == 0, vector_sum_col can be a nullptr
58 if(a_offset != 0)
59 {
60 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
61 ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0));
62 }
63
64 // If b_offset == 0, vector_sum_row can be a nullptr
65 if(b_offset != 0)
66 {
67 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
68
69 // Check if input is a 3D reinterpretation
70 const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
71
72 // Validate input
73 ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2)));
74 ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1));
75
76 TensorShape output_shape = mm_result->tensor_shape();
77 if(output_shape.num_dimensions() > 1)
78 {
79 const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
80
81 TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
82 vector_sum_row_shape.collapse_from(1);
83 output_shape.collapse_from(output_batch_idx);
84
85 ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx],
86 "mm_result tensor must have the same number of batches of output tensor");
87
88 if(a_offset != 0)
89 {
90 TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
91 vector_sum_col_shape.collapse_from(1);
92
93 ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
94 "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
95 }
96 }
97 }
98
99 return Status{};
100 }
101 } // namespace
102
ClGemmLowpOffsetContributionKernel()103 ClGemmLowpOffsetContributionKernel::ClGemmLowpOffsetContributionKernel()
104 {
105 _type = CLKernelType::ELEMENTWISE;
106 }
107
configure(const CLCompileContext & compile_context,const ITensorInfo * mm_result,const ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,const ITensorInfo * bias,int32_t k,int32_t a_offset,int32_t b_offset)108 void ClGemmLowpOffsetContributionKernel::configure(const CLCompileContext &compile_context,
109 const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
110 int32_t k, int32_t a_offset, int32_t b_offset)
111 {
112 // Perform validate step
113 ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result);
114 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, a_offset, b_offset));
115
116 auto padding_info = get_padding_info({ mm_result, vector_sum_col, vector_sum_row, bias });
117
118 // Check if input is a 3D reinterpretation
119 const bool reinterpret_as_3d = vector_sum_row != nullptr
120 && mm_result->num_dimensions() > 1
121 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
122
123 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, mm_result->dimension(0));
124
125 // Set the arguments to pass at compile time
126 CLBuildOptions build_opts;
127 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
128 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mm_result->dimension(0) % num_elems_processed_per_iteration));
129
130 // If a_offset == 0, vector_sum_col can be a nullptr
131 if(a_offset != 0)
132 {
133 build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
134 build_opts.add_option_if(vector_sum_col->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
135 }
136 // If b_offset == 0, vector_sum_row can be a nullptr
137 build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
138 build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * k));
139 build_opts.add_option_if(reinterpret_as_3d, "-DHEIGHT_INPUT3D=" + support::cpp11::to_string(mm_result->dimension(1)));
140 build_opts.add_option_if(reinterpret_as_3d, "-DDEPTH_INPUT3D=" + support::cpp11::to_string(mm_result->dimension(2)));
141 build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
142
143 std::string kernel_name("gemmlowp_offset_contribution");
144
145 // A macro guard to compile ONLY the kernel of interest
146 build_opts.add_option("-D" + upper_string(kernel_name));
147
148 // Create kernel
149 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
150
151 // Configure kernel window
152 Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration));
153 IClKernel::configure_internal(win);
154
155 // Set config_id for enabling LWS tuning
156 _config_id = kernel_name + "_";
157 _config_id += support::cpp11::to_string(mm_result->dimension(0));
158 _config_id += "_";
159 _config_id += support::cpp11::to_string(mm_result->dimension(1));
160 _config_id += "_";
161 _config_id += support::cpp11::to_string(mm_result->dimension(2));
162
163 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
164 }
165
validate(const ITensorInfo * mm_result,const ITensorInfo * vector_sum_col,const ITensorInfo * vector_sum_row,const ITensorInfo * bias,int32_t a_offset,int32_t b_offset)166 Status ClGemmLowpOffsetContributionKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
167 int32_t a_offset, int32_t b_offset)
168 {
169 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, a_offset, b_offset));
170 return Status{};
171 }
172
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)173 void ClGemmLowpOffsetContributionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
174 {
175 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
176 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window);
177
178 const auto vector_sum_col = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_COL_SUM));
179 const auto vector_sum_row = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_ROW_SUM));
180 const auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS));
181 const auto mm_result = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_SRC_DST));
182
183 Window collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ);
184 Window slice = collapsed.first_slice_window_3D();
185
186 // Set window for vector_sum_col
187 Window win_vector_sum_col = slice;
188 win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
189 win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
190
191 // Set window for vector_sum_row
192 Window win_vector_sum_row = slice;
193 win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
194 win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
195 win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
196
197 Window biases_slice = slice;
198 biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
199 biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
200
201 do
202 {
203 unsigned int idx = 0;
204 add_3D_tensor_argument(idx, mm_result, slice);
205 add_2D_tensor_argument_if((vector_sum_col != nullptr), idx, vector_sum_col, win_vector_sum_col);
206 add_2D_tensor_argument_if((vector_sum_row != nullptr), idx, vector_sum_row, win_vector_sum_row);
207 add_1D_tensor_argument_if((bias != nullptr), idx, bias, biases_slice);
208
209 enqueue(queue, *this, slice, lws_hint());
210 }
211 while(collapsed.slide_window_slice_3D(slice));
212 }
213 } // namespace kernels
214 } // namespace opencl
215 } // namespace arm_compute
216