blob: 5c8bf97f0f00ed1e6ff95acfad6834e7e6f88b7f [file] [log] [blame]
Michele Di Giorgio56dd7262017-07-27 09:53:49 +01001/*
Matthew Bentham314d3e22023-06-23 10:53:52 +00002 * Copyright (c) 2017-2021, 2023 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 */
Georgios Pinitas7891a732021-08-20 21:39:25 +010024#include "src/gpu/cl/kernels/ClQuantizeKernel.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"
Manuel Bottini5a1bf622021-03-01 17:39:36 +000029#include "arm_compute/core/Error.h"
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010030#include "arm_compute/core/TensorInfo.h"
31#include "arm_compute/core/Utils.h"
Matthew Bentham314d3e22023-06-23 10:53:52 +000032#include "arm_compute/core/utils/StringUtils.h"
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010033#include "arm_compute/core/Validate.h"
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010034#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
Georgios Pinitasef516e82021-04-30 14:46:05 +010035
Sang-Hoon Park68dd25f2020-10-19 16:00:11 +010036#include "src/core/CL/CLValidate.h"
37#include "src/core/helpers/WindowHelpers.h"
Georgios Pinitasef516e82021-04-30 14:46:05 +010038
Manuel Bottini5a1bf622021-03-01 17:39:36 +000039#include "support/Cast.h"
Matthew Bentham758b5ba2020-03-05 23:37:48 +000040#include "support/StringSupport.h"
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010041
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010042namespace arm_compute
43{
Manuel Bottini5a1bf622021-03-01 17:39:36 +000044namespace opencl
45{
46namespace kernels
47{
Alex Gilday60954c62018-03-05 16:22:48 +000048namespace
49{
Manuel Bottini5a1bf622021-03-01 17:39:36 +000050Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
Alex Gilday60954c62018-03-05 16:22:48 +000051{
Manuel Bottini5a1bf622021-03-01 17:39:36 +000052 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
53 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
54 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +010055
56 // Output must always be initialized
Manuel Bottini5a1bf622021-03-01 17:39:36 +000057 ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0);
58 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QASYMM16);
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
Alex Gilday60954c62018-03-05 16:22:48 +000060
61 return Status{};
62}
Alex Gilday60954c62018-03-05 16:22:48 +000063} // namespace
64
Giorgio Arena4a95bba2021-06-28 11:00:27 +010065ClQuantizeKernel::ClQuantizeKernel()
66{
67 _type = CLKernelType::ELEMENTWISE;
68}
69
Georgios Pinitasef516e82021-04-30 14:46:05 +010070void ClQuantizeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst)
Michele Di Giorgio56dd7262017-07-27 09:53:49 +010071{
Manuel Bottini5a1bf622021-03-01 17:39:36 +000072 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
Manuel Bottini4c6bd512020-04-08 10:15:51 +010073
Manuel Bottini5a1bf622021-03-01 17:39:36 +000074 auto padding_info = get_padding_info({ src, dst });
Sheri Zhang11d73272020-10-28 14:01:55 +000075
Manuel Bottini5a1bf622021-03-01 17:39:36 +000076 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
Sheri Zhang11d73272020-10-28 14:01:55 +000077
Manuel Bottini5a1bf622021-03-01 17:39:36 +000078 const int vec_size_x = 16 / src->element_size();
79 const int input_width_x = src->tensor_shape().x();
Usama Arife03802e2019-03-11 12:20:20 +000080 const bool multi_access_x = (input_width_x / vec_size_x > 0);
81
Manuel Bottini5a1bf622021-03-01 17:39:36 +000082 const UniformQuantizationInfo qinfo = dst->quantization_info().uniform();
83 const DataType output_data_type = dst->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;
Manuel Bottini5a1bf622021-03-01 17:39:36 +000087 if(is_data_type_quantized_asymmetric(src->data_type()))
Manuel Bottini2f602212020-01-30 17:30:32 +000088 {
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 */
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000119 const UniformQuantizationInfo qinfo_in = src->quantization_info().uniform();
Manuel Bottini2f602212020-01-30 17:30:32 +0000120 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 Bottini5a1bf622021-03-01 17:39:36 +0000129 build_opts.add_option_if(is_data_type_float(src->data_type()), "-DIS_FLOAT");
Manuel Bottini2f602212020-01-30 17:30:32 +0000130 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));
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000133 build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(src->data_type()));
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +0100134 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
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000143 Window win = calculate_max_window(*src, Steps());
Sheri Zhang11d73272020-10-28 14:01:55 +0000144 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
Sheri Zhang11d73272020-10-28 14:01:55 +0000150 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
Alex Gilday60954c62018-03-05 16:22:48 +0000151}
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100152
Georgios Pinitasef516e82021-04-30 14:46:05 +0100153Status ClQuantizeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
Alex Gilday60954c62018-03-05 16:22:48 +0000154{
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000155 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
Alex Gilday60954c62018-03-05 16:22:48 +0000156 return Status{};
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100157}
158
Georgios Pinitasef516e82021-04-30 14:46:05 +0100159void ClQuantizeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100160{
161 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
162 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
163
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000164 auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
165 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
166
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100167 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
168 Window slice = window_collapsed.first_slice_window_3D();
169
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100170 do
171 {
172 unsigned int idx = 0;
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000173 add_3D_tensor_argument(idx, src, slice);
174 add_3D_tensor_argument(idx, dst, slice);
Georgios Pinitas275f99c2019-08-23 12:44:11 +0100175 enqueue(queue, *this, slice, lws_hint());
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100176 }
Michalis Spyrou995f5522018-01-29 13:43:35 +0000177 while(window_collapsed.slide_window_slice_3D(slice));
Michele Di Giorgio56dd7262017-07-27 09:53:49 +0100178}
Manuel Bottini5a1bf622021-03-01 17:39:36 +0000179} // namespace kernels
180} // namespace opencl
Michele Di Giorgiod87a7b22019-09-10 10:42:27 +0100181} // namespace arm_compute