1 /*
2 * Copyright (c) 2018-2020 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/CLWidthConcatenate4TensorsKernel.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/Helpers.h"
30 #include "arm_compute/core/Utils.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/WindowHelpers.h"
33 #include "src/core/utils/helpers/tensor_info.h"
34 #include "support/Cast.h"
35
36 #include "support/StringSupport.h"
37
38 namespace arm_compute
39 {
40 namespace
41 {
validate_arguments(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * input3,const ITensorInfo * input4,const ITensorInfo * output)42 Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output)
43 {
44 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, input3, input4, output);
45 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1);
46 ARM_COMPUTE_RETURN_ERROR_ON(input1->data_type() == DataType::UNKNOWN);
47 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2, input3, input4, output);
48 ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) + input2->dimension(0) + input3->dimension(0) + input4->dimension(0) > output->dimension(0));
49
50 for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i)
51 {
52 ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(i) != output->dimension(i));
53 ARM_COMPUTE_RETURN_ERROR_ON(input2->dimension(i) != output->dimension(i));
54 ARM_COMPUTE_RETURN_ERROR_ON(input3->dimension(i) != output->dimension(i));
55 ARM_COMPUTE_RETURN_ERROR_ON(input4->dimension(i) != output->dimension(i));
56 }
57 ARM_COMPUTE_RETURN_ERROR_ON(input1->num_dimensions() > 4);
58
59 return Status{};
60 }
61 } // namespace
62
CLWidthConcatenate4TensorsKernel()63 CLWidthConcatenate4TensorsKernel::CLWidthConcatenate4TensorsKernel()
64 {
65 }
66
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * input3,const ITensorInfo * input4,const ITensorInfo * output)67 Status CLWidthConcatenate4TensorsKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output)
68 {
69 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, input3, input4, output));
70 return Status{};
71 }
72
configure(const CLCompileContext & compile_context,ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * input3,ITensorInfo * input4,ITensorInfo * output)73 void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile_context,
74 ITensorInfo *input1, ITensorInfo *input2,
75 ITensorInfo *input3, ITensorInfo *input4,
76 ITensorInfo *output)
77 {
78 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, input3, input4, output);
79 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, input3, input4, output));
80
81 auto padding_info = get_padding_info({ input1, input2, input3, input4, output });
82 const unsigned int min_dimension = std::min(std::min(input1->dimension(0), input2->dimension(0)), std::min(input3->dimension(0), input4->dimension(0)));
83 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension);
84 const unsigned int vec_size_leftover = output->dimension(0) % num_elems_processed_per_iteration;
85
86 // Add build options
87 CLBuildOptions build_opts;
88 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->data_type()));
89 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
90 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover));
91 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->dimension(2)));
92 build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->dimension(0)));
93 build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->dimension(0)));
94 build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(input3->dimension(0)));
95 build_opts.add_option("-DINPUT4_WIDTH=" + support::cpp11::to_string(input4->dimension(0)));
96 build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->element_size()));
97 build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
98 build_opts.add_option("-DINPUT2_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) + input2->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
99 build_opts.add_option("-DINPUT3_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) + input2->dimension(0) + input3->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
100
101 // If input have different quantization info set quantization parameters needed for the re-quantization process
102 const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output, input1, input2, input3, input4);
103 if(is_data_type_quantized_asymmetric(input1->data_type()) && have_different_qinfo)
104 {
105 const UniformQuantizationInfo iq1_info = input1->quantization_info().uniform();
106 const UniformQuantizationInfo iq2_info = input2->quantization_info().uniform();
107 const UniformQuantizationInfo iq3_info = input3->quantization_info().uniform();
108 const UniformQuantizationInfo iq4_info = input4->quantization_info().uniform();
109 const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
110
111 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq1_info.offset));
112 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale));
113 build_opts.add_option("-DOFFSET_IN2=" + float_to_string_with_full_precision(iq2_info.offset));
114 build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale));
115 build_opts.add_option("-DOFFSET_IN3=" + float_to_string_with_full_precision(iq3_info.offset));
116 build_opts.add_option("-DSCALE_IN3=" + float_to_string_with_full_precision(iq3_info.scale));
117 build_opts.add_option("-DOFFSET_IN4=" + float_to_string_with_full_precision(iq4_info.offset));
118 build_opts.add_option("-DSCALE_IN4=" + float_to_string_with_full_precision(iq4_info.scale));
119 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
120 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
121 }
122
123 // Create kernel
124 _kernel = create_kernel(compile_context, "concatenate_width_x4", build_opts.options());
125
126 // Configure kernel window
127 Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
128 ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
129
130 // Set output valid region
131 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
132 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
133
134 // Set config_id for enabling LWS tuning
135 _config_id = "concatenate_width_x4_";
136 _config_id += lower_string(string_from_data_type(input1->data_type()));
137 _config_id += "_";
138 _config_id += support::cpp11::to_string(input1->dimension(0));
139 _config_id += "_";
140 _config_id += support::cpp11::to_string(input1->dimension(1));
141 _config_id += "_";
142 _config_id += support::cpp11::to_string(input2->dimension(0));
143 _config_id += "_";
144 _config_id += support::cpp11::to_string(input2->dimension(1));
145 _config_id += "_";
146 _config_id += support::cpp11::to_string(input3->dimension(0));
147 _config_id += "_";
148 _config_id += support::cpp11::to_string(input3->dimension(1));
149 _config_id += "_";
150 _config_id += support::cpp11::to_string(input4->dimension(0));
151 _config_id += "_";
152 _config_id += support::cpp11::to_string(input4->dimension(1));
153 }
154
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)155 void CLWidthConcatenate4TensorsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
156 {
157 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
158 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
159
160 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC));
161 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 1));
162 const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 2));
163 const auto src3 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 3));
164 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
165
166 Window slice = window.first_slice_window_4D();
167
168 do
169 {
170 unsigned int idx = 0;
171 add_4D_tensor_argument(idx, src0, slice);
172 add_4D_tensor_argument(idx, src1, slice);
173 add_4D_tensor_argument(idx, src2, slice);
174 add_4D_tensor_argument(idx, src3, slice);
175 add_4D_tensor_argument(idx, dst, slice);
176 enqueue(queue, *this, window, lws_hint());
177 }
178 while(window.slide_window_slice_4D(slice));
179 }
180 } // namespace arm_compute
181