1 /*
2 * Copyright (c) 2019-2022 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/ClHeightConcatenateKernel.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/Helpers.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "src/core/CL/CLValidate.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/Cast.h"
35
36 #include "support/StringSupport.h"
37
38 namespace arm_compute
39 {
40 namespace opencl
41 {
42 namespace kernels
43 {
44 namespace
45 {
validate_arguments(const ITensorInfo * src,unsigned int height_offset,const ITensorInfo * dst)46 Status validate_arguments(const ITensorInfo *src, unsigned int height_offset, const ITensorInfo *dst)
47 {
48 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
49 ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
50 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
51 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) + height_offset > dst->dimension(Window::DimY));
52
53 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != dst->dimension(0));
54 for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(i) != dst->dimension(i));
57 }
58 ARM_COMPUTE_RETURN_ERROR_ON(src->num_dimensions() > 4);
59
60 return Status{};
61 }
62 } // namespace
63
ClHeightConcatenateKernel()64 ClHeightConcatenateKernel::ClHeightConcatenateKernel()
65 : _height_offset(0)
66 {
67 _type = CLKernelType::ELEMENTWISE;
68 }
69
validate(const ITensorInfo * src,unsigned int height_offset,const ITensorInfo * dst)70 Status ClHeightConcatenateKernel::validate(const ITensorInfo *src, unsigned int height_offset, const ITensorInfo *dst)
71 {
72 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, height_offset, dst));
73 return Status{};
74 }
75
configure(const CLCompileContext & compile_context,ITensorInfo * src,unsigned int height_offset,ITensorInfo * dst)76 void ClHeightConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int height_offset, ITensorInfo *dst)
77 {
78 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
79 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, height_offset, dst));
80
81 auto padding_info = get_padding_info({ src, dst });
82
83 _height_offset = height_offset;
84
85 // Add build options
86 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0));
87
88 CLBuildOptions build_opts;
89 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
90 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
91 build_opts.add_option("-DHEIGHT_OFFSET=" + support::cpp11::to_string(_height_offset));
92 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
93
94 if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info())
95 {
96 const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
97 const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
98
99 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
100 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
101 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
102 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
103 }
104 _depth = src->dimension(2);
105
106 std::string kernel_name = "concatenate_height";
107
108 // A macro guard to compile ONLY the kernel of interest
109 build_opts.add_option("-D" + upper_string(kernel_name));
110
111 // Create kernel
112 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
113 // Configure kernel window
114
115 // The window needs to be based on src as we copy all the heights of src
116 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
117 ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
118
119 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
120 }
121
run_op(ITensorPack & tensors,const Window & window,::cl::CommandQueue & queue)122 void ClHeightConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue)
123 {
124 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
125 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
126
127 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
128 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
129
130 unsigned int idx = 0;
131 add_4D_tensor_argument(idx, src, window);
132 add_4D_tensor_argument(idx, dst, window);
133 _kernel.setArg<cl_uint>(idx++, _depth);
134 enqueue(queue, *this, window, lws_hint());
135 }
136 } // namespace kernels
137 } // namespace opencl
138 } // namespace arm_compute
139