1 /*
2 * Copyright (c) 2017-2021 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/gpu/cl/kernels/ClQuantizeKernel.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/Error.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
34
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 #include "support/Cast.h"
39 #include "support/StringSupport.h"
40
41 namespace arm_compute
42 {
43 namespace opencl
44 {
45 namespace kernels
46 {
47 namespace
48 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst)49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
52 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
53 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
54
55 // Output must always be initialized
56 ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0);
57 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QASYMM16);
58 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
59
60 return Status{};
61 }
62 } // namespace
63
ClQuantizeKernel()64 ClQuantizeKernel::ClQuantizeKernel()
65 {
66 _type = CLKernelType::ELEMENTWISE;
67 }
68
configure(const CLCompileContext & compile_context,const ITensorInfo * src,ITensorInfo * dst)69 void ClQuantizeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst)
70 {
71 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
72
73 auto padding_info = get_padding_info({ src, dst });
74
75 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
76
77 const int vec_size_x = 16 / src->element_size();
78 const int input_width_x = src->tensor_shape().x();
79 const bool multi_access_x = (input_width_x / vec_size_x > 0);
80
81 const UniformQuantizationInfo qinfo = dst->quantization_info().uniform();
82 const DataType output_data_type = dst->data_type();
83
84 float scale_to_apply = qinfo.scale;
85 int32_t offset_to_apply = qinfo.offset;
86 if(is_data_type_quantized_asymmetric(src->data_type()))
87 {
88 /*
89 * In case of requantization of a quantized input tensor to an output tensor with another quantization
90 * instead of of apply dequantization and then a quantization functions, we just compute new scale and
91 * offset to apply.
92 *
93 * Assuming:
94 * - q_i as input quantized value
95 * - q_o as output quantized value
96 * - z_i as input quantization offset value
97 * - z_o as output quantization offset value
98 * - s_i as input quantization scale value
99 * - s_o as output quantization scale value
100 * - z_n as new quantization offset value
101 * - s_n as new quantization scale value
102 *
103 * q_o = ( q_i - z_i ) * s_i / s_o + z_o
104 *
105 * We can rewrite the formula as:
106 *
107 * q_o = ( q_i * s_i / s_o ) - z_i * s_i / s_o + z_o
108 *
109 * q_o = q_i / s_n + z_n
110 *
111 * Where:
112 *
113 * s_n = s_o / s_i
114 *
115 * z_n = - z_i * s_i / s_o + z_o
116 *
117 */
118 const UniformQuantizationInfo qinfo_in = src->quantization_info().uniform();
119 scale_to_apply /= qinfo_in.scale;
120 // In order to minimize flooring we convert the offset to a float,
121 // then compute the new offset in the float domain,
122 // finally we convert it back as int32_t
123 offset_to_apply -= static_cast<int32_t>(static_cast<float>(qinfo_in.offset) * qinfo_in.scale / qinfo.scale);
124 }
125
126 // Create kernel
127 CLBuildOptions build_opts;
128 build_opts.add_option_if(is_data_type_float(src->data_type()), "-DIS_FLOAT");
129 build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(scale_to_apply));
130 build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(offset_to_apply));
131 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
132 build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(src->data_type()));
133 build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output_data_type));
134 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)));
135 std::pair<int, int> min_max_quant_values = quantization::get_min_max_values_from_quantized_data_type(output_data_type);
136 build_opts.add_option("-DMIN_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.first));
137 build_opts.add_option("-DMAX_QUANT_VAL=" + support::cpp11::to_string(min_max_quant_values.second));
138
139 _kernel = create_kernel(compile_context, "quantization_layer", build_opts.options());
140
141 // Configure kernel window
142 Window win = calculate_max_window(*src, Steps());
143 if(multi_access_x)
144 {
145 win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
146 }
147 ICLKernel::configure_internal(win);
148
149 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
150 }
151
validate(const ITensorInfo * src,const ITensorInfo * dst)152 Status ClQuantizeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
153 {
154 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
155 return Status{};
156 }
157
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)158 void ClQuantizeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
159 {
160 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
161 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
162
163 auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
164 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
165
166 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
167 Window slice = window_collapsed.first_slice_window_3D();
168
169 do
170 {
171 unsigned int idx = 0;
172 add_3D_tensor_argument(idx, src, slice);
173 add_3D_tensor_argument(idx, dst, slice);
174 enqueue(queue, *this, slice, lws_hint());
175 }
176 while(window_collapsed.slide_window_slice_3D(slice));
177 }
178 } // namespace kernels
179 } // namespace opencl
180 } // namespace arm_compute
181