| 1 |
/* |
| 2 |
* Copyright (c) 2022-2024 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 "depthwise.hpp" |
| 28 |
#include "utils.hpp" |
| 29 |
|
| 30 |
namespace arm_conv { |
| 31 |
namespace depthwise { |
| 32 |
|
| 33 |
template <typename T> struct DefaultTAccum { using Type = T; }; |
| 34 |
template <> struct DefaultTAccum<int8_t> { using Type = int32_t; }; |
| 35 |
template <> struct DefaultTAccum<uint8_t> { using Type = int32_t; }; |
| 36 |
|
| 37 |
template <typename T> struct DefaultOutputStage { using Type = Nothing; }; |
| 38 |
template <> struct DefaultOutputStage<int8_t> { using Type = arm_gemm::Requantize32; }; |
| 39 |
template <> struct DefaultOutputStage<uint8_t> { using Type = arm_gemm::Requantize32; }; |
| 40 |
|
| 41 |
class IDepthfirstStrategy |
| 42 |
{ |
| 43 |
public: |
| 44 |
virtual ~IDepthfirstStrategy() = default; |
| 45 |
|
| 46 |
virtual unsigned int get_input_rows() const = 0; |
| 47 |
virtual unsigned int get_input_cols() const = 0; |
| 48 |
|
| 49 |
virtual unsigned int get_output_rows() const = 0; |
| 50 |
virtual unsigned int get_output_cols() const = 0; |
| 51 |
}; |
| 52 |
|
| 53 |
|
| 54 |
template <typename T> |
| 55 |
struct TensorSpec |
| 56 |
{ |
| 57 |
T base; |
| 58 |
size_t ld_row, ld_col; |
| 59 |
|
| 60 |
TensorSpec(T ptr, size_t ld_row, size_t ld_col) |
| 61 |
: base(ptr), ld_row(ld_row), ld_col(ld_col) {} |
| 62 |
}; |
| 63 |
|
| 64 |
|
| 65 |
template <typename TInput, typename TWeight, typename TOutput> |
| 66 |
class DepthfirstDriver : public DepthwiseCommon<TInput, TWeight, TOutput> |
| 67 |
{ |
| 68 |
protected: |
| 69 |
using Parent = DepthwiseCommon<TInput, TWeight, TOutput>; |
| 70 |
|
| 71 |
// The strategy which we're applying to solve the depthwise convolution. |
| 72 |
std::unique_ptr<const IDepthfirstStrategy> m_strat; |
| 73 |
|
| 74 |
/* Compute the amount of working space required for a single thread. */ |
| 75 |
virtual size_t get_working_size_per_thread() const = 0; |
| 76 |
|
| 77 |
/* Initialise the working space for a thread. */ |
| 78 |
virtual void initialise_working_space(void *) const = 0; |
| 79 |
|
| 80 |
/* Compute a portion of the output tensor with padding. */ |
| 81 |
virtual void compute_tile_padded( |
| 82 |
const DepthwiseArgs &args, |
| 83 |
unsigned int output_i, unsigned int output_j, |
| 84 |
unsigned int output_channel_start, unsigned int output_channel_end, |
| 85 |
const TensorSpec<const TInput *> &input, |
| 86 |
const TensorSpec<TOutput *> &output, |
| 87 |
const void *parameters, |
| 88 |
void *working_space |
| 89 |
) const = 0; |
| 90 |
|
| 91 |
/* Compute a portion of the work with only top/bottom padding. |
| 92 |
* |
| 93 |
* The default implementation of this repeatedly calls into the padded tile |
| 94 |
* variant. |
| 95 |
*/ |
| 96 |
virtual void compute_row_padded_tile_row( |
| 97 |
const DepthwiseArgs &args, |
| 98 |
const unsigned int output_i, unsigned int output_j, unsigned int n_tile_cols, |
| 99 |
const unsigned int output_channel_start, const unsigned int output_channel_end, |
| 100 |
const TensorSpec<const TInput *> &input, |
| 101 |
const TensorSpec<TOutput *> &output, |
| 102 |
const void *parameters, |
| 103 |
void *working_space |
| 104 |
) const |
| 105 |
{ |
| 106 |
for (; n_tile_cols; n_tile_cols--, output_j += m_strat->get_output_cols()) |
| 107 |
{ |
| 108 |
this->compute_tile_padded( |
| 109 |
args, |
| 110 |
output_i, output_j, output_channel_start, output_channel_end, |
| 111 |
input, output, parameters, working_space |
| 112 |
); |
| 113 |
} |
| 114 |
} |
| 115 |
|
| 116 |
/* Compute a portion of the output tensor with no padding. |
| 117 |
* |
| 118 |
* The default implementation of this repeatedly calls into the padded |
| 119 |
* variant. |
| 120 |
*/ |
| 121 |
virtual void compute_tiles_unpadded( |
| 122 |
const DepthwiseArgs &args, |
| 123 |
unsigned int start_output_i, unsigned int start_output_j, |
| 124 |
unsigned int n_tile_rows, unsigned int n_tile_cols, |
| 125 |
unsigned int output_channel_start, unsigned int output_channel_end, |
| 126 |
const TensorSpec<const TInput *> &input, |
| 127 |
const TensorSpec<TOutput *> &output, |
| 128 |
const void *parameters, |
| 129 |
void *working_space |
| 130 |
) const |
| 131 |
{ |
| 132 |
for (unsigned int tile_i = 0; tile_i < n_tile_rows; tile_i++) |
| 133 |
{ |
| 134 |
unsigned int row_start_output_j = start_output_j; |
| 135 |
for (unsigned int tile_j = 0; tile_j < n_tile_cols; tile_j++) |
| 136 |
{ |
| 137 |
this->compute_tile_padded( |
| 138 |
args, |
| 139 |
start_output_i, row_start_output_j, |
| 140 |
output_channel_start, output_channel_end, |
| 141 |
input, output, parameters, working_space |
| 142 |
); |
| 143 |
row_start_output_j += m_strat->get_output_cols(); |
| 144 |
} |
| 145 |
start_output_i += m_strat->get_output_rows(); |
| 146 |
} |
| 147 |
} |
| 148 |
|
| 149 |
void execute_internal( |
| 150 |
const DepthwiseArgs &args, |
| 151 |
const void *input, |
| 152 |
size_t ld_input_col, |
| 153 |
size_t ld_input_row, |
| 154 |
size_t ld_input_batch, |
| 155 |
const void *parameters, |
| 156 |
void *output, |
| 157 |
size_t ld_output_col, |
| 158 |
size_t ld_output_row, |
| 159 |
size_t ld_output_batch, |
| 160 |
void *working_space, |
| 161 |
unsigned int thread_id, |
| 162 |
unsigned int n_threads |
| 163 |
) const override |
| 164 |
{ |
| 165 |
// Get and initialise the working space for this thread. |
| 166 |
void *thread_working_space = |
| 167 |
static_cast<uint8_t *>(working_space) + thread_id * this->get_working_size_per_thread(); |
| 168 |
this->initialise_working_space(thread_working_space); |
| 169 |
|
| 170 |
// Construct convenient representations of the input/output tensors. |
| 171 |
TensorSpec<const TInput *> input_tensor(reinterpret_cast<const TInput *>(input), ld_input_row, ld_input_col); |
| 172 |
TensorSpec<TOutput *> output_tensor(reinterpret_cast<TOutput *>(output), ld_output_row, ld_output_col); |
| 173 |
|
| 174 |
const auto n_output_channels = args.input_channels * args.channel_multiplier; |
| 175 |
|
| 176 |
// By default we parallelize over the rows, but if there's only 1 row, we |
| 177 |
// try to parallize over batches |
| 178 |
auto thread_id_for_rows = thread_id; |
| 179 |
auto n_threads_for_rows = n_threads; |
| 180 |
auto thread_id_for_batches = 0; |
| 181 |
auto n_threads_for_batches = 1; |
| 182 |
if (args.output_rows == 1) { |
| 183 |
thread_id_for_rows = 0; |
| 184 |
n_threads_for_rows = 1; |
| 185 |
thread_id_for_batches = thread_id; |
| 186 |
n_threads_for_batches = n_threads; |
| 187 |
} |
| 188 |
|
| 189 |
// Progress the pointers for the first batch. |
| 190 |
input_tensor.base += ld_input_batch*thread_id_for_batches; |
| 191 |
output_tensor.base += ld_output_batch*thread_id_for_batches; |
| 192 |
for (unsigned int batch = thread_id_for_batches; |
| 193 |
batch < args.n_batches; |
| 194 |
batch += n_threads_for_batches) |
| 195 |
{ |
| 196 |
// Iterate over rows of the output tensor; we stripe over the tiles. |
| 197 |
for (unsigned int start_output_i = thread_id_for_rows * m_strat->get_output_rows(); |
| 198 |
start_output_i < args.output_rows; |
| 199 |
start_output_i += n_threads_for_rows * m_strat->get_output_rows()) |
| 200 |
{ |
| 201 |
// Determine what (if any padding) is required on the top/bottom of |
| 202 |
// this row of the convolution. |
| 203 |
const auto end_output_i = start_output_i + m_strat->get_output_rows(); |
| 204 |
const bool pad_output_bottom = args.output_rows < end_output_i; |
| 205 |
|
| 206 |
const int start_input_i = start_output_i * args.stride_rows - args.padding.top; |
| 207 |
const bool pad_input_top = start_input_i < 0; |
| 208 |
const int end_input_i = start_input_i + m_strat->get_input_rows(); |
| 209 |
const bool pad_input_bottom = static_cast<int>(args.input_rows) < end_input_i; |
| 210 |
// We only need to account for input padding if direct padding is not supported. |
| 211 |
const bool pad_row = ((pad_input_top || pad_input_bottom) && !this->supports_direct_padding()) |
| 212 |
|| pad_output_bottom; |
| 213 |
|
| 214 |
// Iterate over the columns of the output tensor; we attempt to grab as |
| 215 |
// much as possible of the unpadded regions, so the loop structure is a |
| 216 |
// bit odd. |
| 217 |
unsigned int start_output_j = 0; |
| 218 |
while (start_output_j < args.output_cols) |
| 219 |
{ |
| 220 |
const int start_in_j = start_output_j * args.stride_cols - args.padding.left; |
| 221 |
const bool pad_input_left = start_in_j < 0; |
| 222 |
|
| 223 |
// Determine if we can process a number of unpadded tiles in one go. |
| 224 |
int n_unpadded_tiles = 0; |
| 225 |
if ((!pad_input_left) || this->supports_direct_padding()) |
| 226 |
{ |
| 227 |
// Determine the maximum number of tiles we could handle. |
| 228 |
n_unpadded_tiles = (args.output_cols - start_output_j) / m_strat->get_output_cols(); |
| 229 |
|
| 230 |
// Handle padding on the right hand edge |
| 231 |
const int tile_stride = m_strat->get_output_cols() * args.stride_cols; |
| 232 |
int end_output_j = start_output_j + n_unpadded_tiles * m_strat->get_output_cols(); |
| 233 |
int end_input_j = start_in_j + m_strat->get_input_cols() + (n_unpadded_tiles - 1)*tile_stride; |
| 234 |
|
| 235 |
while (n_unpadded_tiles > 0 && |
| 236 |
(static_cast<int>(args.output_cols) < end_output_j || |
| 237 |
static_cast<int>(args.input_cols) < end_input_j)) |
| 238 |
{ |
| 239 |
n_unpadded_tiles--; |
| 240 |
end_output_j -= m_strat->get_output_cols(); |
| 241 |
end_input_j -= tile_stride; |
| 242 |
} |
| 243 |
} |
| 244 |
|
| 245 |
// Process unpadded tiles, if possible, otherwise process a padded tile. |
| 246 |
if (n_unpadded_tiles) |
| 247 |
{ |
| 248 |
if (!pad_row) |
| 249 |
{ |
| 250 |
// Completely unpadded execution |
| 251 |
this->compute_tiles_unpadded( |
| 252 |
args, |
| 253 |
start_output_i, start_output_j, |
| 254 |
1, n_unpadded_tiles, // Compute a row of unpadded tiles |
| 255 |
0, n_output_channels, // Compute all channels |
| 256 |
input_tensor, output_tensor, parameters, thread_working_space |
| 257 |
); |
| 258 |
} |
| 259 |
else |
| 260 |
{ |
| 261 |
// Top/bottom padding only |
| 262 |
this->compute_row_padded_tile_row( |
| 263 |
args, |
| 264 |
start_output_i, start_output_j, n_unpadded_tiles, |
| 265 |
0, n_output_channels, // Compute all channels |
| 266 |
input_tensor, output_tensor, parameters, thread_working_space |
| 267 |
); |
| 268 |
} |
| 269 |
start_output_j += n_unpadded_tiles * m_strat->get_output_cols(); |
| 270 |
} |
| 271 |
else |
| 272 |
{ |
| 273 |
this->compute_tile_padded( |
| 274 |
args, |
| 275 |
start_output_i, start_output_j, |
| 276 |
0, n_output_channels, // Compute all channels |
| 277 |
input_tensor, output_tensor, parameters, thread_working_space |
| 278 |
); |
| 279 |
start_output_j += m_strat->get_output_cols(); |
| 280 |
} |
| 281 |
} |
| 282 |
} |
| 283 |
|
| 284 |
// Progress the pointers for the next batch. |
| 285 |
input_tensor.base += ld_input_batch*n_threads_for_batches; |
| 286 |
output_tensor.base += ld_output_batch*n_threads_for_batches; |
| 287 |
} |
| 288 |
} |
| 289 |
|
| 290 |
public: |
| 291 |
DepthfirstDriver(IDepthfirstStrategy *strategy, const DepthwiseArgs &args) |
| 292 |
: Parent(args), m_strat(strategy) |
| 293 |
{ |
| 294 |
} |
| 295 |
|
| 296 |
size_t get_working_size(unsigned int n_threads) const override final |
| 297 |
{ |
| 298 |
return n_threads * this->get_working_size_per_thread(); |
| 299 |
} |
| 300 |
|
| 301 |
virtual bool supports_direct_padding() const |
| 302 |
{ |
| 303 |
return false; |
| 304 |
} |
| 305 |
}; |
| 306 |
|
| 307 |
} // namespace depthwise |
| 308 |
} // namespace arm_conv |