blob: 76e703f0dd01ffb37fcf579c518a510fc77dd340 [file] [log] [blame]
Michele Di Giorgio56dd7262017-07-27 09:53:49 +01001/*
Michele Di Giorgiod9eaf612020-07-08 11:12:57 +01002 * Copyright (c) 2017-2020 Arm Limited.
Michele Di Giorgio56dd7262017-07-27 09:53:49 +01003 *
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 */
Sang-Hoon Parkbef7fa22020-10-21 15:58:54 +010024#include "src/core/CL/kernels/CLQuantizationLayerKernel.h"
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010025
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010026#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/TensorInfo.h"
30#include "arm_compute/core/Utils.h"
31#include "arm_compute/core/Validate.h"
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010032#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
Sang-Hoon Park68dd25f2020-10-19 16:00:11 +010033#include "src/core/AccessWindowStatic.h"
34#include "src/core/CL/CLValidate.h"
35#include "src/core/helpers/WindowHelpers.h"
Matthew Bentham758b5ba2020-03-05 23:37:48 +000036#include "support/StringSupport.h"
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010037
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010038namespace arm_compute
39{
Alex Gilday60954c62018-03-05 16:22:48 +000040namespace
41{
Usama Arife03802e2019-03-11 12:20:20 +000042Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
Alex Gilday60954c62018-03-05 16:22:48 +000043{
Usama Arife03802e2019-03-11 12:20:20 +000044 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
Manuel Bottini2f602212020-01-30 17:30:32 +000045 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
Usama Arife03802e2019-03-11 12:20:20 +000046 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010047
48 // Output must always be initialized
49 ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape().total_size() == 0);
Manuel Bottini8481d832019-12-10 15:28:40 +000050 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QASYMM16);
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010051 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
Alex Gilday60954c62018-03-05 16:22:48 +000052
53 return Status{};
54}
Alex Gilday60954c62018-03-05 16:22:48 +000055} // namespace
56
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010057CLQuantizationLayerKernel::CLQuantizationLayerKernel()
Usama Arife03802e2019-03-11 12:20:20 +000058 : _input(nullptr), _output(nullptr)
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010059{
60}
61
Usama Arife03802e2019-03-11 12:20:20 +000062void CLQuantizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output)
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010063{
Manuel Bottini4c6bd512020-04-08 10:15:51 +010064 configure(CLKernelLibrary::get().get_compile_context(), input, output);
65}
66
Manuel Bottini679fc962020-04-21 16:08:53 +010067void CLQuantizationLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
Manuel Bottini4c6bd512020-04-08 10:15:51 +010068{
Usama Arife03802e2019-03-11 12:20:20 +000069 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
Sheri Zhang11d73272020-10-28 14:01:55 +000070
71 auto padding_info = get_padding_info({ input, output });
72
Usama Arife03802e2019-03-11 12:20:20 +000073 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010074
Usama Arife03802e2019-03-11 12:20:20 +000075 _input = input;
76 _output = output;
77
78 const int vec_size_x = 16 / input->info()->element_size();
79 const int input_width_x = input->info()->tensor_shape().x();
80 const bool multi_access_x = (input_width_x / vec_size_x > 0);
81
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010082 const UniformQuantizationInfo qinfo = output->info()->quantization_info().uniform();
83 const DataType output_data_type = output->info()->data_type();
Georgios Pinitas4c5469b2019-05-21 13:32:43 +010084
Manuel Bottini2f602212020-01-30 17:30:32 +000085 float scale_to_apply = qinfo.scale;
86 int32_t offset_to_apply = qinfo.offset;
87 if(is_data_type_quantized_asymmetric(_input->info()->data_type()))
88 {
89 /*
90 * In case of requantization of a quantized input tensor to an output tensor with another quantization
91 * instead of of apply dequantization and then a quantization functions, we just compute new scale and
92 * offset to apply.
93 *
94 * Assuming:
95 * - q_i as input quantized value
96 * - q_o as output quantized value
97 * - z_i as input quantization offset value
98 * - z_o as output quantization offset value
99 * - s_i as input quantization scale value
100 * - s_o as output quantization scale value
101 * - z_n as new quantization offset value
102 * - s_n as new quantization scale value
103 *
104 * q_o = ( q_i - z_i ) * s_i / s_o + z_o
105 *
106 * We can rewrite the formula as:
107 *
108 * q_o = ( q_i * s_i / s_o ) - z_i * s_i / s_o + z_o
109 *
110 * q_o = q_i / s_n + z_n
111 *
112 * Where:
113 *
114 * s_n = s_o / s_i
115 *
116 * z_n = - z_i * s_i / s_o + z_o
117 *
118 */
119 const UniformQuantizationInfo qinfo_in = _input->info()->quantization_info().uniform();
120 scale_to_apply /= qinfo_in.scale;
121 // In order to minimize flooring we convert the offset to a float,
122 // then compute the new offset in the float domain,
123 // finally we convert it back as int32_t
124 offset_to_apply -= static_cast<int32_t>(static_cast<float>(qinfo_in.offset) * qinfo_in.scale / qinfo.scale);
125 }
126
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100127 // Create kernel
Usama Arife03802e2019-03-11 12:20:20 +0000128 CLBuildOptions build_opts;
Manuel Bottini2f602212020-01-30 17:30:32 +0000129 build_opts.add_option_if(is_data_type_float(_input->info()->data_type()), "-DIS_FLOAT");
130 build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(scale_to_apply));
131 build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(offset_to_apply));
Usama Arife03802e2019-03-11 12:20:20 +0000132 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +0100133 build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
134 build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output_data_type));
Usama Arife03802e2019-03-11 12:20:20 +0000135 build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(input_width_x - vec_size_x, 0)));
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +0100136 std::pair<int, int> min_max_quant_values = quantization::get_min_max_values_from_quantized_data_type(output_data_type);
137 build_opts.add_option("-DMIN_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.first));
138 build_opts.add_option("-DMAX_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.second));
139
Manuel Bottini4c6bd512020-04-08 10:15:51 +0100140 _kernel = create_kernel(compile_context, "quantization_layer", build_opts.options());
Sheri Zhang11d73272020-10-28 14:01:55 +0000141
142 // Configure kernel window
143 Window win = calculate_max_window(*input->info(), Steps());
144 if(multi_access_x)
145 {
146 win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
147 }
148 ICLKernel::configure_internal(win);
149
150 output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
151
152 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
Alex Gilday60954c62018-03-05 16:22:48 +0000153}
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100154
Usama Arife03802e2019-03-11 12:20:20 +0000155Status CLQuantizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
Alex Gilday60954c62018-03-05 16:22:48 +0000156{
Usama Arife03802e2019-03-11 12:20:20 +0000157 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
Alex Gilday60954c62018-03-05 16:22:48 +0000158 return Status{};
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100159}
160
161void CLQuantizationLayerKernel::run(const Window &window, cl::CommandQueue &queue)
162{
163 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
164 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
165
166 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
167 Window slice = window_collapsed.first_slice_window_3D();
168
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100169 do
170 {
171 unsigned int idx = 0;
172 add_3D_tensor_argument(idx, _input, slice);
173 add_3D_tensor_argument(idx, _output, slice);
Georgios Pinitas275f99c2019-08-23 12:44:11 +0100174 enqueue(queue, *this, slice, lws_hint());
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100175 }
Michalis Spyrou995f5522018-01-29 13:43:35 +0000176 while(window_collapsed.slide_window_slice_3D(slice));
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100177}
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +0100178} // namespace arm_compute