1 /*
2 * Copyright (c) 2019-2021 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/cpu/kernels/CpuConcatenateBatchKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/Window.h"
33 #include "src/core/NEON/NEAsymm.h"
34 #include "src/core/NEON/wrapper/wrapper.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 namespace arm_compute
39 {
40 namespace cpu
41 {
42 namespace kernels
43 {
44 namespace
45 {
46 template <typename T>
batch_concat(const ITensor * src,ITensor * dst,unsigned int batch_offset,const Window & window)47 void batch_concat(const ITensor *src, ITensor *dst, unsigned int batch_offset, const Window &window)
48 {
49 // Offset src
50 uint8_t *src_ptr = src->buffer() + src->info()->offset_first_element_in_bytes();
51
52 // Offset dst
53 uint8_t *dst_ptr = dst->buffer() + dst->info()->offset_first_element_in_bytes() + batch_offset * dst->info()->strides_in_bytes()[3];
54
55 const auto window_start_x = static_cast<int>(window.x().start());
56 const auto window_end_x = static_cast<int>(window.x().end());
57 const int window_step_x = 16 / dst->info()->element_size();
58
59 Window win{ window };
60 win.set(Window::DimX, Window::Dimension(0, 1, 1));
61 win.set(3, Window::Dimension(0, src->info()->tensor_shape()[3], 1));
62
63 Iterator src_it(src, win);
64 Iterator dst_it(dst, win);
65
66 const DataType dt = src->info()->data_type();
67 const UniformQuantizationInfo src_qinfo = src->info()->quantization_info().uniform();
68 const UniformQuantizationInfo dst_qinfo = dst->info()->quantization_info().uniform();
69 if(dt == DataType::QASYMM8 && src_qinfo != dst_qinfo)
70 {
71 execute_window_loop(win, [&](const Coordinates &)
72 {
73 const auto in_ptr = reinterpret_cast<const uint8_t *>(src_ptr + src_it.offset());
74 const auto out_ptr = reinterpret_cast<uint8_t *>(dst_ptr + dst_it.offset());
75
76 int x = window_start_x;
77 for(; x <= (window_end_x - window_step_x); x += window_step_x)
78 {
79 wrapper::vstore(out_ptr, vquantize(vdequantize(wrapper::vloadq(in_ptr), src_qinfo), dst_qinfo));
80 }
81
82 // Compute left-over elements
83 for(; x < window_end_x; ++x)
84 {
85 *(out_ptr + x) = quantize_qasymm8(dequantize_qasymm8(*(in_ptr + x), src_qinfo), dst_qinfo);
86 }
87 },
88 src_it, dst_it);
89 }
90 else if(dt == DataType::QASYMM8_SIGNED && src_qinfo != dst_qinfo)
91 {
92 execute_window_loop(win, [&](const Coordinates &)
93 {
94 const auto in_ptr = reinterpret_cast<const int8_t *>(src_ptr + src_it.offset());
95 const auto out_ptr = reinterpret_cast<int8_t *>(dst_ptr + dst_it.offset());
96 int x = window_start_x;
97 for(; x <= (window_end_x - window_step_x); x += window_step_x)
98 {
99 wrapper::vstore(out_ptr, vquantize_signed(vdequantize(wrapper::vloadq(in_ptr), src_qinfo), dst_qinfo));
100 }
101 // Compute left-over elements
102 for(; x < window_end_x; ++x)
103 {
104 *(out_ptr + x) = quantize_qasymm8_signed(dequantize_qasymm8_signed(*(in_ptr + x), src_qinfo), dst_qinfo);
105 }
106 },
107 src_it, dst_it);
108 }
109 else
110 {
111 execute_window_loop(win, [&](const Coordinates &)
112 {
113 const auto in_ptr = reinterpret_cast<const T *>(src_ptr + src_it.offset());
114 const auto out_ptr = reinterpret_cast<T *>(dst_ptr + dst_it.offset());
115
116 int x = window_start_x;
117 for(; x <= (window_end_x - window_step_x); x += window_step_x)
118 {
119 wrapper::vstore(out_ptr + x, wrapper::vloadq(in_ptr + x));
120 }
121
122 // Compute left-over elements
123 for(; x < window_end_x; ++x)
124 {
125 *(out_ptr + x) = *(in_ptr + x);
126 }
127 },
128 src_it, dst_it);
129 }
130 }
131
validate_arguments(const ITensorInfo * src,unsigned int batch_offset,const ITensorInfo * dst)132 Status validate_arguments(const ITensorInfo *src, unsigned int batch_offset, const ITensorInfo *dst)
133 {
134 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
135 //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src) is not needed here as this kernel doesn't use CPU FP16 instructions.
136 ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
137 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
138
139 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimX) != dst->dimension(Window::DimX));
140 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) != dst->dimension(Window::DimY));
141 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimZ) != dst->dimension(Window::DimZ));
142 ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(3) + batch_offset > dst->dimension(3));
143 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(4, src, dst);
144
145 return Status{};
146 }
147 } // namespace
148
configure(const ITensorInfo * src,unsigned int batch_offset,ITensorInfo * dst)149 void CpuConcatenateBatchKernel::configure(const ITensorInfo *src, unsigned int batch_offset, ITensorInfo *dst)
150 {
151 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
152 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, batch_offset, dst));
153
154 _func = nullptr;
155 _batch_offset = batch_offset;
156
157 switch(src->data_type())
158 {
159 case DataType::S8:
160 case DataType::U8:
161 case DataType::QASYMM8:
162 case DataType::QASYMM8_SIGNED:
163 _func = &batch_concat<uint8_t>;
164 break;
165 case DataType::S16:
166 case DataType::U16:
167 case DataType::F16:
168 _func = &batch_concat<uint16_t>;
169 break;
170 case DataType::S32:
171 case DataType::U32:
172 case DataType::F32:
173 _func = &batch_concat<uint32_t>;
174 break;
175 default:
176 ARM_COMPUTE_ERROR("Unsupported data type.");
177 }
178
179 // Configure kernel window
180 Window win = calculate_max_window(*dst, Steps());
181 ICpuKernel::configure(win);
182 }
183
validate(const arm_compute::ITensorInfo * src,unsigned int batch_offset,const arm_compute::ITensorInfo * dst)184 Status CpuConcatenateBatchKernel::validate(const arm_compute::ITensorInfo *src,
185 unsigned int batch_offset,
186 const arm_compute::ITensorInfo *dst)
187 {
188 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, batch_offset, dst));
189 return Status{};
190 }
191
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)192 void CpuConcatenateBatchKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
193 {
194 ARM_COMPUTE_UNUSED(info);
195 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
196 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
197 ARM_COMPUTE_ERROR_ON(_func == nullptr);
198
199 (*_func)(tensors.get_const_tensor(TensorType::ACL_SRC),
200 tensors.get_tensor(TensorType::ACL_DST),
201 _batch_offset,
202 window);
203 }
204
name() const205 const char *CpuConcatenateBatchKernel::name() const
206 {
207 return "CpuConcatenateBatchKernel";
208 }
209 } // namespace kernels
210 } // namespace cpu
211 } // namespace arm_compute
212