1 /*
2 * Copyright (c) 2018-2020 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/core/CL/kernels/CLStridedSliceKernel.h"
25 #include "arm_compute/core/CL/ICLTensor.h"
26 #include "arm_compute/core/TensorInfo.h"
27 #include "arm_compute/core/utils/helpers/tensor_transform.h"
28 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
29 #include "src/core/helpers/AutoConfiguration.h"
30 #include "src/core/helpers/WindowHelpers.h"
31 #include "src/core/utils/helpers/bit_ops.h"
32 #include "support/Cast.h"
33 #include "support/StringSupport.h"
34
35 namespace arm_compute
36 {
37 namespace
38 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const Coordinates & starts,const Coordinates & ends,const BiStrides & strides,int32_t begin_mask,int32_t end_mask,int32_t shrink_axis_mask)39 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
40 const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
41 int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
42 {
43 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
44 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
45
46 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().num_dimensions() > 4);
47 ARM_COMPUTE_RETURN_ERROR_ON(starts.num_dimensions() > input->num_dimensions());
48 ARM_COMPUTE_RETURN_ERROR_ON(ends.num_dimensions() > input->num_dimensions());
49 ARM_COMPUTE_RETURN_ERROR_ON(strides.num_dimensions() > input->num_dimensions());
50 ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(strides.cbegin(), strides.cbegin() + strides.num_dimensions(), [](int i)
51 {
52 return i == 0;
53 }));
54
55 // Get expected output shape
56 const TensorShape exp_output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
57 starts, ends, strides,
58 begin_mask, end_mask, shrink_axis_mask);
59 ARM_COMPUTE_RETURN_ERROR_ON(exp_output_shape.total_size() == 0);
60
61 // Checks output if configured
62 if(output->total_size() != 0)
63 {
64 const TensorInfo exp_output_info = output->clone()->set_tensor_shape(exp_output_shape);
65 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &exp_output_info);
66 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
67 }
68
69 return Status{};
70 }
71
validate_and_configure_window(const ITensorInfo * input,ITensorInfo * output,const Coordinates & starts,const Coordinates & ends,const BiStrides & strides,int32_t begin_mask,int32_t end_mask,int32_t shrink_axis_mask)72 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input, ITensorInfo *output,
73 const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
74 int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
75 {
76 // Output tensor auto initialization if not yet initialized
77 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
78 starts, ends, strides,
79 begin_mask, end_mask, shrink_axis_mask);
80 auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
81
82 // Create window
83 Window win = calculate_max_window(*output, Steps());
84 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
85
86 return std::make_pair(Status{}, win);
87 }
88 } // namespace
89
configure(const CLCompileContext & compile_context,const ITensorInfo * input,ITensorInfo * output,const Coordinates & starts,const Coordinates & ends,const BiStrides & strides,int32_t begin_mask,int32_t end_mask,int32_t shrink_axis_mask)90 void CLStridedSliceKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *input, ITensorInfo *output,
91 const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
92 int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
93 {
94 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
95 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
96
97 const TensorShape &input_shape = input->tensor_shape();
98
99 Coordinates starts_abs;
100 Coordinates ends_abs;
101 Coordinates final_strides;
102 std::tie(starts_abs, ends_abs, final_strides) = arm_compute::helpers::tensor_transform::calculate_strided_slice_coords(
103 input_shape,
104 starts, ends, strides,
105 begin_mask, end_mask, shrink_axis_mask);
106
107 // Configure kernel window
108 auto win_config = validate_and_configure_window(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask);
109 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
110
111 // Enable multiple elements processing along x if stride_x is 1 and output width greater than the access vector size
112 const int vec_size_x = 16 / input->element_size();
113 const int output_width_x = output->tensor_shape().x();
114 const bool is_shrink_on_x = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, 0);
115 const bool multi_access_x = !is_shrink_on_x && (final_strides.x() == 1) && (output_width_x / vec_size_x > 0);
116
117 // Update window if needed
118 if(multi_access_x)
119 {
120 Window &updated_window = std::get<1>(win_config);
121 updated_window.set(Window::DimX,
122 Window::Dimension(updated_window.x().start(), ceil_to_multiple(updated_window.x().end(), vec_size_x), vec_size_x));
123 }
124 ICLKernel::configure_internal(win_config.second);
125
126 // Create build options
127 CLBuildOptions build_opts;
128 build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(data_size_from_type(input->data_type())));
129 for(unsigned int i = 0; i < input_shape.num_dimensions(); ++i)
130 {
131 const bool is_shrink = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, i);
132 build_opts.add_option("-DSTART_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(starts_abs[i]));
133 build_opts.add_option("-DSTRIDE_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(final_strides[i]));
134 build_opts.add_option_if(is_shrink, "-DSHRINK_" + support::cpp11::to_string(i));
135 }
136 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)));
137 build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
138 build_opts.add_option_if_else(input_shape.num_dimensions() > 2,
139 "-DSRC_DEPTH=" + support::cpp11::to_string(input_shape.z()),
140 "-DSRC_DEPTH=1");
141 build_opts.add_option_if_else(output->num_dimensions() > 2,
142 "-DDST_DEPTH=" + support::cpp11::to_string(output->tensor_shape().z()),
143 "-DDST_DEPTH=1");
144
145 // Create kernel
146 _kernel = create_kernel(compile_context, "strided_slice", build_opts.options());
147
148 // Set config_id for enabling LWS tuning
149 _config_id = "strided_slice";
150 _config_id += "_";
151 _config_id += lower_string(string_from_data_type(input->data_type()));
152 for(unsigned int i = 0; i < input_shape.num_dimensions(); ++i)
153 {
154 _config_id += "_";
155 _config_id += support::cpp11::to_string(input->dimension(i));
156 _config_id += "_";
157 _config_id += support::cpp11::to_string(starts_abs[i]);
158 _config_id += "_";
159 _config_id += support::cpp11::to_string(ends_abs[i]);
160 _config_id += "_";
161 _config_id += support::cpp11::to_string(final_strides[i]);
162 }
163 }
164
validate(const ITensorInfo * input,const ITensorInfo * output,const Coordinates & starts,const Coordinates & ends,const BiStrides & strides,int32_t begin_mask,int32_t end_mask,int32_t shrink_axis_mask)165 Status CLStridedSliceKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
166 const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
167 int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
168 {
169 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
170 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(),
171 starts, ends, strides, begin_mask, end_mask, shrink_axis_mask)
172 .first);
173
174 return Status{};
175 }
176
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)177 void CLStridedSliceKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
178 {
179 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
180 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
181
182 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
183 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
184
185 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
186 Window slice = window_collapsed.first_slice_window_4D();
187
188 do
189 {
190 unsigned int idx = 0;
191 add_4D_tensor_argument(idx, src, slice);
192 add_4D_tensor_argument(idx, dst, slice);
193 enqueue(queue, *this, slice, lws_hint());
194 }
195 while(window_collapsed.slide_window_slice_4D(slice));
196 }
197 } // namespace arm_compute
198