| 1 |
/* |
| 2 |
* Copyright (c) 2021-2024 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 |
|
| 25 |
#pragma once |
| 26 |
|
| 27 |
#include "depthfirst_driver.hpp" |
| 28 |
#include "src/core/NEON/kernels/arm_conv/addressing.hpp" |
| 29 |
#include "utils.hpp" |
| 30 |
#if !defined(_WIN64) && !defined(__OpenBSD__) |
| 31 |
#include <alloca.h> |
| 32 |
#endif /* !defined(_WIN64) && !defined(__OpenBSD__) */ |
| 33 |
#include <limits> |
| 34 |
|
| 35 |
namespace arm_conv { |
| 36 |
namespace pooling { |
| 37 |
|
| 38 |
template <typename TInput, typename TOutput> |
| 39 |
class DepthfirstStrategy : public IDepthfirstStrategy |
| 40 |
{ |
| 41 |
unsigned int input_rows, input_cols, output_rows, output_cols; |
| 42 |
|
| 43 |
public: |
| 44 |
DepthfirstStrategy(unsigned int window_rows, unsigned int window_cols, |
| 45 |
unsigned int stride_rows, unsigned int stride_cols, |
| 46 |
unsigned int output_rows, unsigned int output_cols) |
| 47 |
: input_rows(output_rows + (window_rows - 1) * stride_rows), |
| 48 |
input_cols(output_cols + (window_cols - 1) * stride_cols), |
| 49 |
output_rows(output_rows), output_cols(output_cols) |
| 50 |
{ |
| 51 |
} |
| 52 |
|
| 53 |
unsigned int get_input_rows() const override { return input_rows; } |
| 54 |
unsigned int get_input_cols() const override { return input_cols; } |
| 55 |
unsigned int get_output_rows() const override { return output_rows; } |
| 56 |
unsigned int get_output_cols() const override { return output_cols; } |
| 57 |
|
| 58 |
typedef void (*KernelType)( |
| 59 |
unsigned int n_channels, |
| 60 |
const TInput *const *, |
| 61 |
TOutput *const *, |
| 62 |
bool exclude_padding, |
| 63 |
unsigned int pad_left, |
| 64 |
unsigned int pad_top, |
| 65 |
unsigned int pad_right, |
| 66 |
unsigned int pad_bottom |
| 67 |
); |
| 68 |
virtual KernelType get_kernel(void) const = 0; |
| 69 |
}; |
| 70 |
|
| 71 |
|
| 72 |
struct WorkingSpace |
| 73 |
{ |
| 74 |
void *input_buffer; |
| 75 |
void *output_buffer; |
| 76 |
}; |
| 77 |
|
| 78 |
|
| 79 |
template <typename TInput, typename TOutput=TInput, class OutputStage=Nothing> |
| 80 |
class PoolingDepthfirst : public DepthfirstDriver<TInput, TOutput> |
| 81 |
{ |
| 82 |
size_t sizeof_input_buffer(void) const |
| 83 |
{ |
| 84 |
return sizeof(TInput) * this->m_args.n_channels; |
| 85 |
} |
| 86 |
|
| 87 |
size_t sizeof_output_buffer(void) const |
| 88 |
{ |
| 89 |
return sizeof(TOutput) * this->m_args.n_channels; |
| 90 |
} |
| 91 |
|
| 92 |
protected: |
| 93 |
/* Compute the amount of working space required for a single thread. */ |
| 94 |
size_t get_working_size_per_thread(unsigned int n_channels) const override |
| 95 |
{ |
| 96 |
return sizeof(WorkingSpace) + n_channels * (sizeof(TInput) + sizeof(TOutput)); |
| 97 |
} |
| 98 |
|
| 99 |
/* Initialise the working space for a thread. */ |
| 100 |
void initialise_working_space(void *raw_ws, unsigned int n_channels) const override |
| 101 |
{ |
| 102 |
auto ws = reinterpret_cast<WorkingSpace *>(raw_ws); |
| 103 |
ws->input_buffer = ws + 1; |
| 104 |
ws->output_buffer = reinterpret_cast<char *>(ws + 1) + sizeof(TInput) * n_channels; |
| 105 |
|
| 106 |
// Fill the input buffer with an appropriate value |
| 107 |
TInput fill_val = 0; |
| 108 |
if (this->m_args.pool_type == PoolingType::MAX) |
| 109 |
{ |
| 110 |
using limits = std::numeric_limits<TInput>; |
| 111 |
if (limits::has_infinity) |
| 112 |
{ |
| 113 |
fill_val = -limits::infinity(); |
| 114 |
} |
| 115 |
else |
| 116 |
{ |
| 117 |
fill_val = limits::min(); |
| 118 |
} |
| 119 |
} |
| 120 |
|
| 121 |
auto ptr = reinterpret_cast<TInput *>(ws->input_buffer); |
| 122 |
for (; n_channels; n_channels--) |
| 123 |
{ |
| 124 |
*(ptr++) = fill_val; |
| 125 |
} |
| 126 |
} |
| 127 |
|
| 128 |
/* Compute a portion of the output tensor with padding. */ |
| 129 |
void compute_tile_padded( |
| 130 |
unsigned int output_i, unsigned int output_j, |
| 131 |
unsigned int channel_start, unsigned int channel_end, |
| 132 |
const TensorSpec<const TInput *> &input, |
| 133 |
const TensorSpec<TOutput *> &output, |
| 134 |
void *working_space |
| 135 |
) const override |
| 136 |
{ |
| 137 |
const auto kern = reinterpret_cast<const DepthfirstStrategy<TInput, TOutput> *>( |
| 138 |
this->m_strat.get())->get_kernel(); |
| 139 |
|
| 140 |
// Get the working space, and some space on the stack for pointer arrays |
| 141 |
auto ws = reinterpret_cast<WorkingSpace *>(working_space); |
| 142 |
auto inptr_array = reinterpret_cast<const TInput **>(alloca( |
| 143 |
sizeof(TInput *) * this->m_strat->get_input_rows() * this->m_strat->get_input_cols())); |
| 144 |
auto outptr_array = reinterpret_cast<TOutput **>(alloca( |
| 145 |
sizeof(TOutput *) * this->m_strat->get_output_rows() * this->m_strat->get_output_cols())); |
| 146 |
|
| 147 |
// Prepare the input pointers |
| 148 |
const int ii = static_cast<int>(output_i * this->m_args.pool_stride.rows) - this->m_args.padding.top; |
| 149 |
const auto input_pad_top = static_cast<unsigned int>(ii < 0 ? -ii : 0); |
| 150 |
const auto input_i = static_cast<unsigned int>(ii < 0 ? 0 : ii); |
| 151 |
|
| 152 |
const unsigned int end_ii = ii + this->m_strat->get_input_rows(); |
| 153 |
const auto input_pad_bottom = end_ii < this->m_args.input_rows ? 0 : end_ii - this->m_args.input_rows; |
| 154 |
|
| 155 |
const int ij = static_cast<int>(output_j * this->m_args.pool_stride.cols) - this->m_args.padding.left; |
| 156 |
const auto input_pad_left = static_cast<unsigned int>(ij < 0 ? -ij : 0); |
| 157 |
const auto input_j = static_cast<unsigned int>(ij < 0 ? 0 : ij); |
| 158 |
|
| 159 |
const unsigned int end_ij = ij + this->m_strat->get_input_cols(); |
| 160 |
const auto input_pad_right = end_ij < this->m_args.input_cols ? 0 : end_ij - this->m_args.input_cols; |
| 161 |
|
| 162 |
fill_pointer_array<const TInput>( |
| 163 |
inptr_array, this->m_strat->get_input_rows(), this->m_strat->get_input_cols(), |
| 164 |
input.base + input_i*input.ld_row + input_j*input.ld_col + channel_start, |
| 165 |
input.ld_row, input.ld_col, |
| 166 |
reinterpret_cast<const TInput *>(ws->input_buffer), |
| 167 |
input_pad_top, this->m_args.input_rows - input_i, |
| 168 |
input_pad_left, this->m_args.input_cols - input_j |
| 169 |
); |
| 170 |
|
| 171 |
// Prepare the output pointers |
| 172 |
fill_pointer_array( |
| 173 |
outptr_array, this->m_strat->get_output_rows(), this->m_strat->get_output_cols(), |
| 174 |
output.base + output_i*output.ld_row + output_j*output.ld_col + channel_start, |
| 175 |
output.ld_row, output.ld_col, |
| 176 |
reinterpret_cast<TOutput *>(ws->output_buffer), |
| 177 |
0, this->m_args.output_rows - output_i, // Top padding, # valid rows |
| 178 |
0, this->m_args.output_cols - output_j // Left padding, # valid columns |
| 179 |
); |
| 180 |
|
| 181 |
// Call the kernel |
| 182 |
kern( |
| 183 |
channel_end - channel_start, inptr_array, outptr_array, |
| 184 |
this->m_args.exclude_padding, |
| 185 |
input_pad_left, input_pad_top, |
| 186 |
input_pad_right, input_pad_bottom |
| 187 |
); |
| 188 |
} |
| 189 |
|
| 190 |
// Compute a portion of the work with only top/bottom padding. |
| 191 |
void compute_row_padded_tile_row( |
| 192 |
const unsigned int output_i, unsigned int output_j, unsigned int n_tile_cols, |
| 193 |
const unsigned int channel_start, const unsigned int channel_end, |
| 194 |
const TensorSpec<const TInput *> &input, |
| 195 |
const TensorSpec<TOutput *> &output, |
| 196 |
void *working_space |
| 197 |
) const override |
| 198 |
{ |
| 199 |
const auto kern = reinterpret_cast<const DepthfirstStrategy<TInput, TOutput> *>( |
| 200 |
this->m_strat.get())->get_kernel(); |
| 201 |
|
| 202 |
// Get the working space, and some space on the stack for pointer arrays |
| 203 |
auto ws = reinterpret_cast<WorkingSpace *>(working_space); |
| 204 |
auto inptr_array = reinterpret_cast<const TInput **>(alloca( |
| 205 |
sizeof(TInput *) * this->m_strat->get_input_rows() * this->m_strat->get_input_cols())); |
| 206 |
auto outptr_array = reinterpret_cast<TOutput **>(alloca( |
| 207 |
sizeof(TOutput *) * this->m_strat->get_output_rows() * this->m_strat->get_output_cols())); |
| 208 |
|
| 209 |
// Prepare the initial input pointers |
| 210 |
const int ii = static_cast<int>(output_i * this->m_args.pool_stride.rows) - this->m_args.padding.top; |
| 211 |
const auto input_pad_top = static_cast<unsigned int>(ii < 0 ? -ii : 0); |
| 212 |
const auto input_i = static_cast<unsigned int>(ii < 0 ? 0 : ii); |
| 213 |
|
| 214 |
const unsigned int end_ii = ii + this->m_strat->get_input_rows(); |
| 215 |
const auto input_pad_bottom = end_ii < this->m_args.input_rows ? 0 : end_ii - this->m_args.input_rows; |
| 216 |
|
| 217 |
const int ij = static_cast<int>(output_j * this->m_args.pool_stride.cols) - this->m_args.padding.left; |
| 218 |
const auto input_j = static_cast<unsigned int>(ij < 0 ? 0 : ij); |
| 219 |
|
| 220 |
const auto end_oi = output_i + this->m_strat->get_output_cols(); |
| 221 |
const auto output_pad_bottom = end_oi < this->m_args.output_rows ? 0 : end_oi - this->m_args.output_rows; |
| 222 |
|
| 223 |
fill_pointer_array<const TInput>( |
| 224 |
inptr_array, this->m_strat->get_input_rows(), this->m_strat->get_input_cols(), |
| 225 |
input.base + input_i*input.ld_row + input_j*input.ld_col + channel_start, |
| 226 |
input.ld_row, input.ld_col, |
| 227 |
reinterpret_cast<const TInput *>(ws->input_buffer), |
| 228 |
input_pad_top, this->m_args.input_rows - input_i, |
| 229 |
0, this->m_args.input_cols - input_j |
| 230 |
); |
| 231 |
|
| 232 |
// Prepare the initial output pointers |
| 233 |
fill_pointer_array( |
| 234 |
outptr_array, this->m_strat->get_output_rows(), this->m_strat->get_output_cols(), |
| 235 |
output.base + output_i*output.ld_row + output_j*output.ld_col + channel_start, |
| 236 |
output.ld_row, output.ld_col, |
| 237 |
reinterpret_cast<TOutput *>(ws->output_buffer), |
| 238 |
0, this->m_args.output_rows - output_i, // Top padding, # valid rows |
| 239 |
0, this->m_args.output_cols - output_j // Left padding, # valid columns |
| 240 |
); |
| 241 |
|
| 242 |
// Call the kernel |
| 243 |
for (; n_tile_cols; n_tile_cols--) |
| 244 |
{ |
| 245 |
kern( |
| 246 |
channel_end - channel_start, inptr_array, outptr_array, |
| 247 |
this->m_args.exclude_padding, |
| 248 |
0, input_pad_top, |
| 249 |
0, input_pad_bottom |
| 250 |
); |
| 251 |
|
| 252 |
// Progress the input and output pointer arrays |
| 253 |
const auto input_col_stride = input.ld_col * this->m_strat->get_output_cols() * this->m_args.pool_stride.cols; |
| 254 |
for ( |
| 255 |
auto n = input_pad_top * this->m_strat->get_input_cols(); |
| 256 |
n < (this->m_strat->get_input_rows() - input_pad_bottom) * this->m_strat->get_input_cols(); |
| 257 |
n++ |
| 258 |
) |
| 259 |
{ |
| 260 |
inptr_array[n] += input_col_stride; |
| 261 |
} |
| 262 |
|
| 263 |
const auto output_col_stride = output.ld_col * this->m_strat->get_output_cols(); |
| 264 |
for ( |
| 265 |
auto n = 0u; |
| 266 |
n < (this->m_strat->get_output_rows() - output_pad_bottom) * this->m_strat->get_output_cols(); |
| 267 |
n++ |
| 268 |
) |
| 269 |
{ |
| 270 |
outptr_array[n] += output_col_stride; |
| 271 |
} |
| 272 |
} |
| 273 |
} |
| 274 |
|
| 275 |
public: |
| 276 |
PoolingDepthfirst(const DepthfirstStrategy<TInput, TOutput> *strat, |
| 277 |
const PoolingArgs &args, const OutputStage &os = {}) |
| 278 |
: DepthfirstDriver<TInput, TOutput>(strat, args) |
| 279 |
{ |
| 280 |
ARM_COMPUTE_UNUSED(os); |
| 281 |
} |
| 282 |
}; |
| 283 |
|
| 284 |
} // namespace pooling |
| 285 |
} // namespace arm_conv |