• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/CLWidthConcatenate2TensorsKernel.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/Utils.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.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 * output)42 Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
43 {
44     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, 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, output);
48     ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) + input2->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     }
55     ARM_COMPUTE_RETURN_ERROR_ON(input1->num_dimensions() > 4);
56 
57     return Status{};
58 }
59 } // namespace
60 
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output)61 Status CLWidthConcatenate2TensorsKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
62 {
63     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output));
64     return Status{};
65 }
66 
configure(const CLCompileContext & compile_context,ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output)67 void CLWidthConcatenate2TensorsKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
68 {
69     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
70     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, output));
71 
72     auto padding_info = get_padding_info({ input1, input2, output });
73 
74     const unsigned int min_dimension                     = std::min(input1->dimension(0), input2->dimension(0));
75     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension);
76     const unsigned int vec_size_leftover                 = output->dimension(0) % num_elems_processed_per_iteration;
77 
78     // Add build options
79     CLBuildOptions build_opts;
80     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->data_type()));
81     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
82     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover));
83     build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->dimension(2)));
84     build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->dimension(0)));
85     build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->dimension(0)));
86     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->element_size()));
87     build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
88 
89     // If input have different quantization info set quantization parameters needed for the re-quantization process
90     const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output, input1, input2);
91     if(is_data_type_quantized_asymmetric(input1->data_type()) && have_different_qinfo)
92     {
93         const UniformQuantizationInfo iq1_info = input1->quantization_info().uniform();
94         const UniformQuantizationInfo iq2_info = input2->quantization_info().uniform();
95         const UniformQuantizationInfo oq_info  = output->quantization_info().uniform();
96 
97         build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq1_info.offset));
98         build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale));
99         build_opts.add_option("-DOFFSET_IN2=" + float_to_string_with_full_precision(iq2_info.offset));
100         build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale));
101         build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
102         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
103     }
104 
105     // Create kernel
106     _kernel = create_kernel(compile_context, "concatenate_width_x2", build_opts.options());
107 
108     // Configure kernel window
109     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
110     ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
111 
112     // Set output valid region
113     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
114     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
115 
116     // Set config_id for enabling LWS tuning
117     _config_id = "concatenate_width_x2_";
118     _config_id += lower_string(string_from_data_type(input1->data_type()));
119     _config_id += "_";
120     _config_id += support::cpp11::to_string(input1->dimension(0));
121     _config_id += "_";
122     _config_id += support::cpp11::to_string(input1->dimension(1));
123     _config_id += "_";
124     _config_id += support::cpp11::to_string(input2->dimension(0));
125     _config_id += "_";
126     _config_id += support::cpp11::to_string(input2->dimension(1));
127 }
128 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)129 void CLWidthConcatenate2TensorsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
130 {
131     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
132     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
133 
134     Window slice = window.first_slice_window_4D();
135 
136     const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC));
137     const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 1));
138     auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
139 
140     do
141     {
142         unsigned int idx = 0;
143         add_4D_tensor_argument(idx, src0, slice);
144         add_4D_tensor_argument(idx, src1, slice);
145         add_4D_tensor_argument(idx, dst, slice);
146         enqueue(queue, *this, window, lws_hint());
147     }
148     while(window.slide_window_slice_4D(slice));
149 }
150 } // namespace arm_compute
151