xref: /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ClCastKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2016-2022 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/ClCastKernel.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/CL/OpenCL.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
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,ConvertPolicy policy)48 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, ConvertPolicy policy)
49 {
50     ARM_COMPUTE_UNUSED(policy);
51     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
52     ARM_COMPUTE_RETURN_ERROR_ON(src == dst);
53     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src,
54                                                          1,
55                                                          DataType::U8, DataType::S8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::S16,
56                                                          DataType::U16, DataType::U32, DataType::S32, DataType::F16,
57                                                          DataType::F32);
58     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst,
59                                                          1,
60                                                          DataType::U8, DataType::S8, DataType::QASYMM8, DataType::S16,
61                                                          DataType::U16, DataType::U32, DataType::S32, DataType::F16,
62                                                          DataType::F32);
63     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_type() == dst->data_type(), "src and dst data types must be different");
64 
65     // Validate in case of configured dst
66     if(dst->total_size() > 0)
67     {
68         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
69     }
70 
71     return Status{};
72 }
73 } // namespace
74 
ClCastKernel()75 ClCastKernel::ClCastKernel()
76 {
77     _type = CLKernelType::ELEMENTWISE;
78 }
79 
configure(const CLCompileContext & compile_context,const ITensorInfo * src,ITensorInfo * dst,ConvertPolicy policy)80 void ClCastKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, ConvertPolicy policy)
81 {
82     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
83 
84     // Auto initialize dst shape if not initialized (We can only auto-configure the shape, datatype must be given)
85     set_shape_if_empty(*dst, src->tensor_shape());
86 
87     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, policy));
88 
89     auto padding_info = get_padding_info({ src, dst });
90 
91     // Get data sizes
92     const size_t src_size = data_size_from_type(src->data_type());
93     const size_t dst_size = data_size_from_type(dst->data_type());
94 
95     // Get number of elements to process per iterations
96     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0));
97 
98     // Set build options
99     CLBuildOptions build_opts;
100     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
101     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
102     build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(src->data_type()));
103     build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(dst->data_type()));
104     // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
105     build_opts.add_option_if(is_data_type_float(src->data_type()) || policy == ConvertPolicy::SATURATE, "-DSATURATE");
106     build_opts.add_option_if(is_data_type_float(src->data_type()) || is_data_type_float(dst->data_type()), "-DIS_DATA_TYPE_FLOAT");
107     build_opts.add_option_if(is_data_type_quantized(src->data_type()), "-DIS_DATA_TYPE_QUANTIZED");
108 
109     // Create kernel
110     const std::string kernel_name = (src_size >= dst_size) ? "cast_down" : "cast_up";
111     _kernel                       = create_kernel(compile_context, kernel_name, build_opts.options());
112 
113     // Configure kernel
114     Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
115     ICLKernel::configure_internal(win);
116 
117     // Collapse window
118     const Window &full_window      = window();
119     Window        collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ);
120     ICLKernel::configure_internal(collapsed_window);
121 
122     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
123 
124     // Set config_id for enabling LWS tuning
125     _config_id = kernel_name;
126     _config_id += "_";
127     _config_id += lower_string(string_from_data_type(src->data_type()));
128     _config_id += "_";
129     _config_id += support::cpp11::to_string(src->dimension(0));
130     _config_id += "_";
131     _config_id += support::cpp11::to_string(src->dimension(1));
132     _config_id += "_";
133     _config_id += support::cpp11::to_string(dst->dimension(0));
134     _config_id += "_";
135     _config_id += support::cpp11::to_string(dst->dimension(1));
136 }
137 
validate(const ITensorInfo * src,const ITensorInfo * dst,ConvertPolicy policy)138 Status ClCastKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, ConvertPolicy policy)
139 {
140     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, policy));
141     return Status{};
142 }
143 
run_op(ITensorPack & tensors,const Window & window,::cl::CommandQueue & queue)144 void ClCastKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue)
145 {
146     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
147     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
148 
149     const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
150     auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
151 
152     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
153 
154     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
155     Window slice     = collapsed.first_slice_window_3D();
156 
157     do
158     {
159         unsigned int idx = 0;
160         add_3D_tensor_argument(idx, src, slice);
161         add_3D_tensor_argument(idx, dst, slice);
162         enqueue(queue, *this, slice, lws_hint());
163     }
164     while(collapsed.slide_window_slice_3D(slice));
165 }
166 } // namespace kernels
167 } // namespace opencl
168 } // namespace arm_compute
169