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/ClCol2ImKernel.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/Helpers.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "src/core/CL/CLValidate.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 #include "support/Cast.h"
36 #include "support/StringSupport.h"
37
38 #include <cmath>
39
40 namespace arm_compute
41 {
42 using namespace misc::shape_calculator;
43 namespace opencl
44 {
45 namespace kernels
46 {
47 namespace
48 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const Size2D & convolved_dims,unsigned int num_groups)49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
52 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
53 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
54
55 // Checks performed when output is configured
56 if(dst->total_size() != 0)
57 {
58 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), compute_col2im_shape(*src, convolved_dims, true, num_groups));
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
60 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_layout() != DataLayout::NCHW, "Col2Im output's data layout must always be NCHW");
62 }
63
64 return Status{};
65 }
66
validate_and_configure_window(ITensorInfo * src,ITensorInfo * dst,const Size2D & convolved_dims,unsigned int num_groups)67 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
68 {
69 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
70 // Output auto inizialitation if not yet initialized
71 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_col2im_shape(*src, convolved_dims, true, num_groups)).set_data_layout(DataLayout::NCHW));
72
73 constexpr unsigned int num_elems_read_per_iteration = 8;
74
75 // Configure window
76 Window win = calculate_max_window(*src, Steps(num_elems_read_per_iteration));
77
78 // Update window and padding just for the input tensor as we cannot access out-of-bounds elements in the output one
79 AccessWindowHorizontal input_access(src, 0, num_elems_read_per_iteration);
80 bool window_changed = update_window_and_padding(win, input_access);
81
82 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
83 return std::make_pair(err, win);
84 }
85 } // namespace
86
ClCol2ImKernel()87 ClCol2ImKernel::ClCol2ImKernel()
88 : _convolved_dims()
89 {
90 _type = CLKernelType::ELEMENTWISE;
91 }
92
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,const Size2D & convolved_dims,unsigned int num_groups)93 void ClCol2ImKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
94 {
95 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
96
97 // Perform validation step
98 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, convolved_dims, num_groups));
99
100 _convolved_dims = convolved_dims;
101
102 const DataType data_type = src->data_type();
103
104 // Create kernel
105 CLBuildOptions build_opts;
106 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
107 build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src->element_size()));
108 build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(src->dimension(0)));
109 build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.width));
110 build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
111
112 _kernel = create_kernel(compile_context, "col2im", build_opts.options());
113
114 // Configure kernel window
115 auto win_config = validate_and_configure_window(src, dst, _convolved_dims, num_groups);
116 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
117 IClKernel::configure_internal(win_config.second);
118
119 // Set config_id for enabling LWS tuning
120 _config_id = "col2im_";
121 _config_id += lower_string(string_from_data_type(src->data_type()));
122 _config_id += "_";
123 _config_id += support::cpp11::to_string(num_groups);
124 _config_id += "_";
125 _config_id += support::cpp11::to_string(src->dimension(0));
126 _config_id += "_";
127 _config_id += support::cpp11::to_string(src->dimension(1));
128 _config_id += "_";
129 _config_id += support::cpp11::to_string(dst->dimension(0));
130 _config_id += "_";
131 _config_id += support::cpp11::to_string(dst->dimension(1));
132 }
133
validate(const ITensorInfo * src,const ITensorInfo * dst,const Size2D & convolved_dims,unsigned int num_groups)134 Status ClCol2ImKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
135 {
136 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
137 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, convolved_dims, num_groups));
138 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), convolved_dims, num_groups).first);
139 return Status{};
140 }
141
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)142 void ClCol2ImKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
143 {
144 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
145 ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IClKernel::window(), window);
146
147 bool is_collapsed = false;
148 bool is_collapsed_out = false;
149
150 auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
151 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
152
153 Window out_window;
154 out_window.use_tensor_dimensions(dst->info()->tensor_shape());
155
156 Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &is_collapsed);
157 Window collapsed_out = out_window.collapse_if_possible(out_window, 3, &is_collapsed_out);
158
159 ARM_COMPUTE_ERROR_ON(is_collapsed != is_collapsed_out);
160
161 Window slice = collapsed.first_slice_window_3D();
162 Window slice_out = collapsed_out.first_slice_window_4D();
163 do
164 {
165 // Set inputs
166 unsigned int idx = 0;
167 add_3D_tensor_argument(idx, src, slice);
168 add_4D_tensor_argument(idx, dst, slice_out);
169 enqueue(queue, *this, slice, lws_hint());
170 }
171 while(collapsed.slide_window_slice_3D(slice) && collapsed_out.slide_window_slice_4D(slice_out));
172 }
173 } // namespace kernels
174 } // namespace opencl
175 } // namespace arm_compute
176