1 /*
2 * Copyright (c) 2017-2022 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/ClDepthConcatenateKernel.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 "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/WindowHelpers.h"
33 #include "support/Cast.h"
34
35 #include "support/StringSupport.h"
36
37 namespace arm_compute
38 {
39 namespace opencl
40 {
41 namespace kernels
42 {
43 namespace
44 {
validate_arguments(const ITensorInfo * src,unsigned int depth_offset,const ITensorInfo * dst)45 Status validate_arguments(const ITensorInfo *src, unsigned int depth_offset, const ITensorInfo *dst)
46 {
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
48 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
49 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
50 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
51
52 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimX) != dst->dimension(Window::DimX));
53 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) != dst->dimension(Window::DimY));
54 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(2) + depth_offset > dst->dimension(2));
55 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, src, dst);
56
57 return Status{};
58 }
59 } // namespace
60
ClDepthConcatenateKernel()61 ClDepthConcatenateKernel::ClDepthConcatenateKernel()
62 : _depth_offset(0)
63 {
64 _type = CLKernelType::ELEMENTWISE;
65 }
66
configure(const CLCompileContext & compile_context,ITensorInfo * src,unsigned int depth_offset,ITensorInfo * dst)67 void ClDepthConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int depth_offset, ITensorInfo *dst)
68 {
69 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
70 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, depth_offset, dst));
71
72 auto padding_info = get_padding_info({ src, dst });
73
74 _depth_offset = depth_offset;
75
76 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0));
77
78 // Add build options
79 CLBuildOptions build_opts;
80 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
81 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
82 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
83 if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info())
84 {
85 const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
86 const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
87
88 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
89 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
90 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
91 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
92 }
93
94 std::string kernel_name = "concatenate";
95
96 // A macro guard to compile ONLY the kernel of interest
97 build_opts.add_option("-D" + upper_string(kernel_name));
98
99 // Create kernel
100 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
101
102 // Configure kernel window
103 auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration));
104 win.set(Window::DimZ, Window::Dimension(0, src->tensor_shape().z(), 1));
105 ICLKernel::configure_internal(win);
106
107 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
108 }
109
validate(const arm_compute::ITensorInfo * src,unsigned int depth_offset,const arm_compute::ITensorInfo * dst)110 Status ClDepthConcatenateKernel::validate(const arm_compute::ITensorInfo *src,
111 unsigned int depth_offset,
112 const arm_compute::ITensorInfo *dst)
113 {
114 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, depth_offset, dst));
115 return Status{};
116 }
117
run_op(ITensorPack & tensors,const Window & window,::cl::CommandQueue & queue)118 void ClDepthConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue)
119 {
120 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
121 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
122
123 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
124 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
125
126 Window slice = window.first_slice_window_3D();
127
128 const int offset_to_first_elements_in_bytes = _depth_offset * dst->info()->strides_in_bytes()[2];
129
130 unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the src and dst parameters
131 _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes);
132
133 do
134 {
135 unsigned int idx = 0;
136 add_3D_tensor_argument(idx, src, slice);
137 add_3D_tensor_argument(idx, dst, slice);
138 enqueue(queue, *this, slice, lws_hint());
139 }
140 while(window.slide_window_slice_3D(slice));
141 }
142 } // namespace kernels
143 } // namespace opencl
144 } // namespace arm_compute
145