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/ClDequantizeKernel.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/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32
33 #include "src/core/CL/CLValidate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include "support/Cast.h"
38 #include "support/StringSupport.h"
39
40 namespace arm_compute
41 {
42 namespace opencl
43 {
44 namespace kernels
45 {
46 namespace
47 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst)48 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst)
49 {
50 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16);
52
53 if(dst->tensor_shape().total_size() > 0)
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(dst);
56 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::F16, DataType::F32);
57 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
58 }
59
60 return Status{};
61 }
62 } // namespace
63
ClDequantizeKernel()64 ClDequantizeKernel::ClDequantizeKernel()
65 {
66 _type = CLKernelType::ELEMENTWISE;
67 }
68
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst)69 void ClDequantizeKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst)
70 {
71 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
72
73 // Output tensor auto initialization if not yet initialized
74 auto_init_if_empty(*dst, src->tensor_shape(), 1, DataType::F32);
75
76 auto padding_info = get_padding_info({ src, dst });
77
78 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst));
79
80 const int vec_size_x = 16 / dst->element_size();
81 const int output_width_x = dst->tensor_shape().x();
82 const bool multi_access_x = (output_width_x / vec_size_x > 0);
83
84 const bool is_quantized_per_channel = is_data_type_quantized_per_channel(src->data_type());
85 std::string kernel_name = "dequantization_layer";
86
87 // Create kernel
88 CLBuildOptions build_opts;
89 if(!is_quantized_per_channel)
90 {
91 const UniformQuantizationInfo qinfo = src->quantization_info().uniform();
92 const int qoffset = is_data_type_quantized_asymmetric(src->data_type()) ? qinfo.offset : 0;
93 build_opts.add_option("-DSCALE=" + float_to_string_with_full_precision(qinfo.scale));
94 build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(qoffset));
95 }
96 else
97 {
98 kernel_name += "_per_channel";
99 kernel_name += src->data_layout() == DataLayout::NCHW ? "_nchw" : "_nhwc";
100 }
101
102 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
103 build_opts.add_option("-DDATA_TYPE_SRC=" + get_cl_type_from_data_type(src->data_type()));
104 build_opts.add_option("-DDATA_TYPE_DST=" + get_cl_type_from_data_type(dst->data_type()));
105 build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(output_width_x - vec_size_x, 0)));
106
107 // Create kernel name
108 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
109
110 // Configure kernel window
111 Window win = calculate_max_window(*dst);
112 if(multi_access_x)
113 {
114 win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x));
115 }
116 ICLKernel::configure_internal(win);
117
118 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
119 }
120
validate(const ITensorInfo * src,const ITensorInfo * dst)121 Status ClDequantizeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
122 {
123 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst));
124 return Status{};
125 }
126
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)127 void ClDequantizeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
128 {
129 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
130 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
131
132 auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
133 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
134
135 const bool is_quantized_per_channel = is_data_type_quantized_per_channel(src->info()->data_type());
136
137 // Collapse windo
138 Window new_window = is_quantized_per_channel ? window.collapse_if_possible(ICLKernel::window(), 4) : window.collapse_if_possible(ICLKernel::window(), 3);
139 Window slice = new_window.first_slice_window_3D();
140
141 if(is_quantized_per_channel)
142 {
143 unsigned int idx = num_arguments_per_3D_tensor() * 2; //Skip the input and output parameters
144 _kernel.setArg(idx++, src->quantization().scale->cl_buffer());
145 }
146
147 do
148 {
149 unsigned int idx = 0;
150 add_3D_tensor_argument(idx, src, slice);
151 add_3D_tensor_argument(idx, dst, slice);
152 enqueue(queue, *this, slice, lws_hint());
153 }
154 while(new_window.slide_window_slice_3D(slice));
155 }
156 } // namespace kernels
157 } // namespace opencl
158 } // namespace arm_compute
159