• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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/CLWeightsReshapeKernel.h"
25 #include "arm_compute/core/CL/ICLTensor.h"
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "src/core/helpers/AutoConfiguration.h"
29 #include "src/core/helpers/WindowHelpers.h"
30 #include "support/StringSupport.h"
31 
32 namespace arm_compute
33 {
34 using namespace arm_compute::misc::shape_calculator;
35 
36 namespace
37 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * biases,const ITensorInfo * output,unsigned int num_groups)38 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, unsigned int num_groups)
39 {
40     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
41     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
42     ARM_COMPUTE_RETURN_ERROR_ON(num_groups == 0);
43     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::NHWC && num_groups > 1);
44     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4 && num_groups > 1);
45     ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(3) % num_groups) != 0);
46 
47     if(biases != nullptr)
48     {
49         ARM_COMPUTE_RETURN_ERROR_ON(!is_data_type_float(input->data_type()));
50         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
51         ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1));
52         ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2));
53         ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3]));
54         ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->dimension(0) != input->tensor_shape()[3] || biases->dimension(1) != input->tensor_shape()[4]));
55     }
56 
57     // Checks performed when output is configured
58     if(output->total_size() != 0)
59     {
60         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_weights_reshaped_shape(*input, biases != nullptr, num_groups));
61         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
62         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
63     }
64 
65     return Status{};
66 }
67 } // namespace
68 
CLWeightsReshapeKernel()69 CLWeightsReshapeKernel::CLWeightsReshapeKernel()
70     : _input(nullptr), _biases(nullptr), _output(nullptr)
71 {
72 }
73 
configure(const ICLTensor * input,const ICLTensor * biases,ICLTensor * output,unsigned int num_groups)74 void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, unsigned int num_groups)
75 {
76     configure(CLKernelLibrary::get().get_compile_context(), input, biases, output, num_groups);
77 }
78 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * biases,ICLTensor * output,unsigned int num_groups)79 void CLWeightsReshapeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, unsigned int num_groups)
80 {
81     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
82 
83     // Output tensor auto inizialitation if not yet initialized
84     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_weights_reshaped_shape(*input->info(), (biases != nullptr), num_groups)));
85 
86     // Perform validation step
87     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
88                                                   (biases != nullptr) ? biases->info() : nullptr,
89                                                   output->info(), num_groups));
90 
91     auto padding_info = get_padding_info({ input, biases, output });
92 
93     const DataType data_type = input->info()->data_type();
94 
95     _biases = biases;
96     _output = output;
97     _input  = input;
98 
99     // Create build options
100     CLBuildOptions build_opts;
101     build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(data_size_from_type(data_type)));
102     build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
103     build_opts.add_option_if(biases != nullptr, "-DHAS_BIAS");
104 
105     // Create kernel
106     _kernel = create_kernel(compile_context, "reshape_to_columns", build_opts.options());
107 
108     // Configure window
109     Window win = calculate_max_window(*input->info(), Steps());
110     // The CLWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
111     output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
112     ICLKernel::configure_internal(win);
113 
114     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
115 }
116 
validate(const ITensorInfo * input,const ITensorInfo * biases,const ITensorInfo * output,unsigned int num_groups)117 Status CLWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, unsigned int num_groups)
118 {
119     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, biases, output, num_groups));
120     return Status{};
121 }
122 
run(const Window & window,cl::CommandQueue & queue)123 void CLWeightsReshapeKernel::run(const Window &window, cl::CommandQueue &queue)
124 {
125     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
126     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
127 
128     Window out_window;
129     out_window.use_tensor_dimensions(_output->info()->tensor_shape());
130 
131     Window in_slice  = window.first_slice_window_3D();
132     Window out_slice = out_window.first_slice_window_2D();
133 
134     Window biases_window;
135     Window biases_slice;
136 
137     unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
138     idx += (_biases != nullptr) ? num_arguments_per_1D_tensor() : 0;
139     _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(0));
140     _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(1));
141     _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(2));
142     _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(3));
143     _kernel.setArg<cl_uint>(idx++, _output->info()->strides_in_bytes().z());
144 
145     if(_biases != nullptr)
146     {
147         biases_window.use_tensor_dimensions(_biases->info()->tensor_shape());
148         biases_slice = biases_window.first_slice_window_1D();
149     }
150 
151     do
152     {
153         // Set arguments
154         unsigned idx = 0;
155         add_3D_tensor_argument(idx, _input, in_slice);
156         add_2D_tensor_argument(idx, _output, out_slice);
157         if(_biases != nullptr)
158         {
159             add_1D_tensor_argument(idx, _biases, biases_slice);
160             ARM_COMPUTE_UNUSED(biases_window.slide_window_slice_1D(biases_slice));
161         }
162 
163         // Run kernel
164         enqueue(queue, *this, in_slice, lws_hint());
165     }
166     while(window.slide_window_slice_4D(in_slice) && out_window.slide_window_slice_2D(out_slice));
167 }
168 } // namespace arm_compute
169