xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2021 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/cpu/kernels/CpuConvertQuantizedSignednessKernel.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 cpu
39 {
40 namespace kernels
41 {
42 namespace
43 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst)44 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
45 {
46     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
47     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
48 
49     // Validate output if initialized
50     if(dst->total_size() != 0)
51     {
52         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
53         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(src->tensor_shape(), dst->tensor_shape());
54     }
55 
56     return Status{};
57 }
58 
validate_and_configure_window(const ITensorInfo * src,ITensorInfo * dst)59 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src, ITensorInfo *dst)
60 {
61     // Output auto inizialitation if not yet initialized
62     {
63         const bool                    is_input_signed   = src->data_type() == DataType::QASYMM8_SIGNED;
64         const DataType                dt                = is_input_signed ? DataType::QASYMM8 : DataType::QASYMM8_SIGNED;
65         const UniformQuantizationInfo qinfo             = src->quantization_info().uniform();
66         const int                     offset_correction = is_input_signed ? -128 : 128;
67         const QuantizationInfo        corrected_qinfo   = QuantizationInfo(qinfo.scale, qinfo.offset + offset_correction);
68 
69         auto_init_if_empty(*dst, src->clone()->set_data_type(dt).set_quantization_info(corrected_qinfo));
70     }
71 
72     return std::make_pair(Status{}, calculate_max_window(*dst));
73 }
74 } // namespace
75 
configure(const ITensorInfo * src,ITensorInfo * dst)76 void CpuConvertQuantizedSignednessKernel::configure(const ITensorInfo *src, ITensorInfo *dst)
77 {
78     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
79     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
80 
81     std::pair<Status, Window> win_config = validate_and_configure_window(src, dst);
82     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
83     ICpuKernel::configure(win_config.second);
84 }
85 
validate(const ITensorInfo * src,const ITensorInfo * dst)86 Status CpuConvertQuantizedSignednessKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
87 {
88     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
89     return Status{};
90 }
91 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)92 void CpuConvertQuantizedSignednessKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
93 {
94     auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
95     auto dst = tensors.get_tensor(TensorType::ACL_DST);
96     ARM_COMPUTE_UNUSED(info);
97     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
98     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
99 
100     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
101     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
102 
103     Iterator input(src, win_collapsed);
104     Iterator output(dst, win_collapsed);
105 
106     const int  window_step_x  = 16;
107     const auto window_start_x = static_cast<int>(window.x().start());
108     const auto window_end_x   = static_cast<int>(window.x().end());
109 
110     const uint8_t mask  = 128;
111     const auto    vmask = wrapper::vdup_n(mask, wrapper::traits::vector_128_tag{});
112 
113     execute_window_loop(win_collapsed, [&](const Coordinates &)
114     {
115         const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr());
116         const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
117 
118         // Compute S elements per iteration
119         int x = window_start_x;
120         for(; x <= (window_end_x - window_step_x); x += window_step_x)
121         {
122             const auto vin = wrapper::vloadq(input_ptr + x);
123             wrapper::vstore(output_ptr + x, wrapper::veor(vin, vmask));
124         }
125 
126         // Compute left-over elements
127         for(; x < window_end_x; ++x)
128         {
129             const uint8_t in  = *(reinterpret_cast<const uint8_t *>(input_ptr + x));
130             *(output_ptr + x) = in ^ mask;
131         }
132     },
133     input, output);
134 }
135 
name() const136 const char *CpuConvertQuantizedSignednessKernel::name() const
137 {
138     return "CpuConvertQuantizedSignednessKernel";
139 }
140 } // namespace kernels
141 } // namespace cpu
142 } // namespace arm_compute
143