1 /*
2  * Copyright (c) 2018-2022 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/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 
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36 
37 #include "support/Cast.h"
38 #include "support/StringSupport.h"
39 
40 namespace arm_compute
41 {
42 namespace opencl
43 {
44 namespace kernels
45 {
46 namespace
47 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const GEMMLowpOutputStageInfo * info)48 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *info)
49 {
50     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32);
51     ARM_COMPUTE_RETURN_ERROR_ON((info->output_data_type != DataType::QASYMM8) && (info->output_data_type != DataType::QASYMM8_SIGNED));
52     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)));
53     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))
54                                 || info->gemmlowp_min_bound > info->gemmlowp_max_bound);
55 
56     // Check biases if exist
57     if(bias != nullptr)
58     {
59         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias);
60         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
61         ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
62     }
63 
64     if(dst->total_size() != 0)
65     {
66         ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_type() != info->output_data_type, "Mismatching output data type");
67         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
68     }
69 
70     return Status{};
71 }
72 } // namespace
73 
ClGemmLowpQuantizeDownInt32ScaleByFloatKernel()74 ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel()
75 {
76     _type = CLKernelType::ELEMENTWISE;
77 }
78 
validate(const ITensorInfo * src,const ITensorInfo * bias,const ITensorInfo * dst,const GEMMLowpOutputStageInfo * info)79 Status ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst,
80                                                                const GEMMLowpOutputStageInfo *info)
81 {
82     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
83     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, info));
84 
85     return Status{};
86 }
87 
configure(const CLCompileContext & compile_context,const ITensorInfo * src,const ITensorInfo * bias,ITensorInfo * dst,const GEMMLowpOutputStageInfo * info)88 void ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst,
89                                                               const GEMMLowpOutputStageInfo *info)
90 {
91     // Perform validate step
92     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
93     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, info));
94 
95     auto padding_info = get_padding_info({ src, bias, dst });
96 
97     // Output auto inizialitation if not yet initialized
98     auto_init_if_empty(*dst, src->clone()->set_data_type(info->output_data_type));
99 
100     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0));
101 
102     auto min = info->gemmlowp_min_bound;
103     auto max = info->gemmlowp_max_bound;
104 
105     // Set the arguments to pass at compile time
106     CLBuildOptions build_opts;
107     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
108     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
109     build_opts.add_option("-DREAL_MULTIPLIER=" + float_to_string_with_full_precision(info->gemmlowp_real_multiplier));
110     build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(info->gemmlowp_offset));
111     build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type()));
112     build_opts.add_option_if((min > 0), "-DMIN_BOUND=" + support::cpp11::to_string(min));
113     build_opts.add_option_if((max < 255), "-DMAX_BOUND=" + support::cpp11::to_string(max));
114     build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
115 
116     const std::string kernel_name = "gemmlowp_output_stage_quantize_down_float";
117 
118     // A macro guard to compile ONLY the kernel of interest
119     build_opts.add_option("-D" + upper_string(kernel_name));
120 
121     // Create kernel
122     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
123 
124     // Configure kernel window
125     Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
126     ICLKernel::configure_internal(win);
127 
128     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
129 }
130 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)131 void ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
132 {
133     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
134     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
135 
136     const auto src  = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
137     const auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS));
138     auto       dst  = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
139 
140     // Create input window
141     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
142     Window slice     = collapsed.first_slice_window_3D();
143 
144     // Setup bias slice
145     unsigned int idx1 = num_arguments_per_3D_tensor();
146     if(bias != nullptr)
147     {
148         Window biases_slice(slice);
149         biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
150         biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
151         add_1D_tensor_argument(idx1, bias, biases_slice);
152     }
153 
154     do
155     {
156         unsigned int idx = 0;
157         add_3D_tensor_argument(idx, src, slice);
158         add_3D_tensor_argument(idx1, dst, slice);
159         enqueue(queue, *this, slice, lws_hint());
160     }
161     while(collapsed.slide_window_slice_3D(slice));
162 }
163 } // namespace kernels
164 } // namespace opencl
165 } // namespace arm_compute
166