• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/ICLTensor.h"
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
33 #include "src/core/AccessWindowStatic.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36 #include "support/StringSupport.h"
37 
38 namespace arm_compute
39 {
40 namespace
41 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const GEMMLowpOutputStageInfo * info)42 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
43                           const GEMMLowpOutputStageInfo *info)
44 {
45     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
46     ARM_COMPUTE_RETURN_ERROR_ON((info->output_data_type != DataType::QASYMM8) && (info->output_data_type != DataType::QASYMM8_SIGNED));
47     ARM_COMPUTE_RETURN_ERROR_ON(info->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(info->output_data_type)));
48     ARM_COMPUTE_RETURN_ERROR_ON(info->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(info->output_data_type))
49                                 || info->gemmlowp_min_bound > info->gemmlowp_max_bound);
50 
51     // Check biases if exist
52     if(bias != nullptr)
53     {
54         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
55         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
56         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
57     }
58 
59     if(output->total_size() != 0)
60     {
61         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() != info->output_data_type, "Mismatching output data type");
62         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
63     }
64 
65     return Status{};
66 }
67 } // namespace
68 
69 class Coordinates;
CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel()70 CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel()
71     : _input(nullptr), _bias(nullptr), _output(nullptr)
72 {
73 }
74 
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,const GEMMLowpOutputStageInfo * info)75 Status CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
76                                                                const GEMMLowpOutputStageInfo *info)
77 {
78     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
79     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, info));
80 
81     return Status{};
82 }
83 
configure(const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const GEMMLowpOutputStageInfo * info)84 void CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
85                                                               const GEMMLowpOutputStageInfo *info)
86 {
87     configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, info);
88 }
89 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * bias,ICLTensor * output,const GEMMLowpOutputStageInfo * info)90 void CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
91                                                               const GEMMLowpOutputStageInfo *info)
92 {
93     // Perform validate step
94     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
95     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), info));
96 
97     auto padding_info = get_padding_info({ input, bias, output });
98 
99     // Output auto inizialitation if not yet initialized
100     auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(info->output_data_type));
101 
102     _input  = input;
103     _bias   = bias;
104     _output = output;
105 
106     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, input->info()->dimension(0));
107 
108     auto min = info->gemmlowp_min_bound;
109     auto max = info->gemmlowp_max_bound;
110 
111     // Set the arguments to pass at compile time
112     CLBuildOptions build_opts;
113     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
114     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration));
115     build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(info->gemmlowp_real_multiplier));
116     build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(info->gemmlowp_offset));
117     build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
118     build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min));
119     build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max));
120     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
121 
122     // Create kernel
123     _kernel = create_kernel(compile_context, "gemmlowp_output_stage_quantize_down_float", build_opts.options());
124 
125     // Configure kernel window
126     Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
127     ICLKernel::configure_internal(win);
128 
129     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
130 }
131 
run(const Window & window,cl::CommandQueue & queue)132 void CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel::run(const Window &window, cl::CommandQueue &queue)
133 {
134     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
135     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
136 
137     // Create input window
138     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
139     Window slice     = collapsed.first_slice_window_3D();
140 
141     // Setup bias slice
142     unsigned int idx1 = num_arguments_per_3D_tensor();
143     if(_bias != nullptr)
144     {
145         Window biases_slice(slice);
146         biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
147         biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
148         add_1D_tensor_argument(idx1, _bias, biases_slice);
149     }
150 
151     do
152     {
153         unsigned int idx = 0;
154         add_3D_tensor_argument(idx, _input, slice);
155         add_3D_tensor_argument(idx1, _output, slice);
156         enqueue(queue, *this, slice, lws_hint());
157     }
158     while(collapsed.slide_window_slice_3D(slice));
159 }
160 } // namespace arm_compute
161