• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2020 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/NEON/kernels/NEConvertQuantizedSignednessKernel.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/Window.h"
32 #include "src/core/NEON/wrapper/wrapper.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 
36 namespace arm_compute
37 {
38 namespace
39 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output)40 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
41 {
42     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
43     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
44 
45     // Validate output if initialized
46     if(output->total_size() != 0)
47     {
48         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
49         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output->tensor_shape());
50     }
51 
52     return Status{};
53 }
54 
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)55 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
56 {
57     // Output auto inizialitation if not yet initialized
58     {
59         const bool                    is_input_signed   = input->data_type() == DataType::QASYMM8_SIGNED;
60         const DataType                dt                = is_input_signed ? DataType::QASYMM8 : DataType::QASYMM8_SIGNED;
61         const UniformQuantizationInfo qinfo             = input->quantization_info().uniform();
62         const int                     offset_correction = is_input_signed ? -128 : 128;
63         const QuantizationInfo        corrected_qinfo   = QuantizationInfo(qinfo.scale, qinfo.offset + offset_correction);
64 
65         auto_init_if_empty(*output, input->clone()->set_data_type(dt).set_quantization_info(corrected_qinfo));
66     }
67 
68     return std::make_pair(Status{}, calculate_max_window(*output));
69 }
70 } // namespace
71 
NEConvertQuantizedSignednessKernel()72 NEConvertQuantizedSignednessKernel::NEConvertQuantizedSignednessKernel()
73     : _input(nullptr), _output(nullptr)
74 {
75 }
76 
configure(const ITensor * input,ITensor * output)77 void NEConvertQuantizedSignednessKernel::configure(const ITensor *input, ITensor *output)
78 {
79     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
80     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
81 
82     _input  = input;
83     _output = output;
84 
85     std::pair<Status, Window> win_config = validate_and_configure_window(input->info(), output->info());
86     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
87     INEKernel::configure(win_config.second);
88 }
89 
validate(const arm_compute::ITensorInfo * input,const arm_compute::ITensorInfo * output)90 Status NEConvertQuantizedSignednessKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output)
91 {
92     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
93     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
94     return Status{};
95 }
96 
run(const Window & window,const ThreadInfo & info)97 void NEConvertQuantizedSignednessKernel::run(const Window &window, const ThreadInfo &info)
98 {
99     ARM_COMPUTE_UNUSED(info);
100     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
101     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
102 
103     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
104     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
105 
106     Iterator input(_input, win_collapsed);
107     Iterator output(_output, win_collapsed);
108 
109     const int  window_step_x  = 16;
110     const auto window_start_x = static_cast<int>(window.x().start());
111     const auto window_end_x   = static_cast<int>(window.x().end());
112 
113     const uint8_t mask  = 128;
114     const auto    vmask = wrapper::vdup_n(mask, wrapper::traits::vector_128_tag{});
115 
116     execute_window_loop(win_collapsed, [&](const Coordinates &)
117     {
118         const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr());
119         const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
120 
121         // Compute S elements per iteration
122         int x = window_start_x;
123         for(; x <= (window_end_x - window_step_x); x += window_step_x)
124         {
125             const auto vin = wrapper::vloadq(input_ptr + x);
126             wrapper::vstore(output_ptr + x, wrapper::veor(vin, vmask));
127         }
128 
129         // Compute left-over elements
130         for(; x < window_end_x; ++x)
131         {
132             const uint8_t in  = *(reinterpret_cast<const uint8_t *>(input_ptr + x));
133             *(output_ptr + x) = in ^ mask;
134         }
135     },
136     input, output);
137 }
138 } // namespace arm_compute
139