| 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 "depthfirst_driver.hpp" |
| 28 |
#include "interleaves/generic.hpp" |
| 29 |
|
| 30 |
namespace arm_conv { |
| 31 |
namespace depthwise { |
| 32 |
|
| 33 |
template <typename OutputStage> |
| 34 |
class IPlanarStrategy |
| 35 |
{ |
| 36 |
public: |
| 37 |
virtual ~IPlanarStrategy() = default; |
| 38 |
virtual unsigned int get_output_rows(void) const = 0; |
| 39 |
virtual arm_gemm::VLType get_vl_type(void) const = 0; |
| 40 |
|
| 41 |
virtual size_t get_storage_size(const DepthwiseArgs &) const = 0; |
| 42 |
virtual void pack_parameters( |
| 43 |
const DepthwiseArgs &args, void *buffer, |
| 44 |
const void *biases, const OutputStage &, |
| 45 |
const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 46 |
) const = 0; |
| 47 |
}; |
| 48 |
|
| 49 |
|
| 50 |
template <typename TInput, typename TWeight, typename TOutput, typename TAccum, |
| 51 |
typename OutputStage> |
| 52 |
struct PlanarKernelType; |
| 53 |
|
| 54 |
template <typename TInput, typename TWeight, typename TOutput, typename TAccum> |
| 55 |
struct PlanarKernelType<TInput, TWeight, TOutput, TAccum, Nothing> |
| 56 |
{ |
| 57 |
typedef void (*Type)( |
| 58 |
const TInput *, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 59 |
unsigned int pad_top, unsigned int valid_input_rows, |
| 60 |
unsigned int pad_left, unsigned int valid_input_cols, |
| 61 |
const TWeight *, const TAccum *, |
| 62 |
TOutput **, const size_t *, const size_t *, unsigned int output_cols, |
| 63 |
unsigned int start_channels, unsigned int valid_channels, |
| 64 |
TAccum act_min, TAccum act_max |
| 65 |
); |
| 66 |
|
| 67 |
template <typename WorkspaceType> |
| 68 |
static inline void execute( |
| 69 |
const Type fn, |
| 70 |
const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 71 |
unsigned int pad_top, unsigned int valid_input_rows, |
| 72 |
unsigned int pad_left, unsigned int valid_input_cols, |
| 73 |
const TWeight *weights, const TAccum *bias, |
| 74 |
TOutput **outptrs, const size_t *outlds, const size_t *outvllds, unsigned int output_cols, |
| 75 |
unsigned int start_channel, unsigned int valid_channels, |
| 76 |
const Nothing &, const WorkspaceType *ws |
| 77 |
) |
| 78 |
{ |
| 79 |
fn( |
| 80 |
inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 81 |
pad_top, valid_input_rows, |
| 82 |
pad_left, valid_input_cols, |
| 83 |
weights, bias, |
| 84 |
outptrs, outlds, outvllds, output_cols, |
| 85 |
start_channel, valid_channels, |
| 86 |
ws->activation_min, ws->activation_max |
| 87 |
); |
| 88 |
} |
| 89 |
}; |
| 90 |
|
| 91 |
template <typename TInput, typename TWeight, typename TOutput> |
| 92 |
struct PlanarKernelType<TInput, TWeight, TOutput, int32_t, arm_gemm::Requantize32> |
| 93 |
{ |
| 94 |
typedef void (*Type)( |
| 95 |
const TInput *, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 96 |
unsigned int pad_top, unsigned int valid_input_rows, |
| 97 |
unsigned int pad_left, unsigned int valid_input_cols, |
| 98 |
const TWeight *, |
| 99 |
TOutput **, const size_t *, const size_t *, unsigned int output_cols, |
| 100 |
unsigned int start_channel, unsigned int valid_channels, |
| 101 |
const arm_gemm::Requantize32 & |
| 102 |
); |
| 103 |
|
| 104 |
template <typename WorkspaceType> |
| 105 |
static inline void execute( |
| 106 |
const Type fn, |
| 107 |
const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 108 |
unsigned int pad_top, unsigned int valid_input_rows, |
| 109 |
unsigned int pad_left, unsigned int valid_input_cols, |
| 110 |
const TWeight *weights, const int32_t *, |
| 111 |
TOutput **outptrs, const size_t *outlds, const size_t *outldvls, unsigned int output_cols, |
| 112 |
unsigned int first_channel, unsigned int valid_channels, |
| 113 |
const arm_gemm::Requantize32 &qp, const WorkspaceType * |
| 114 |
) |
| 115 |
{ |
| 116 |
fn( |
| 117 |
inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 118 |
pad_top, valid_input_rows, |
| 119 |
pad_left, valid_input_cols, |
| 120 |
weights, |
| 121 |
outptrs, outlds, outldvls, output_cols, |
| 122 |
first_channel, valid_channels, |
| 123 |
qp |
| 124 |
); |
| 125 |
} |
| 126 |
}; |
| 127 |
|
| 128 |
|
| 129 |
template <typename TInput, typename TWeight=TInput, typename TOutput=TInput, |
| 130 |
typename TAccum=typename DefaultTAccum<TOutput>::Type, |
| 131 |
typename OutputStage=typename DefaultOutputStage<TOutput>::Type> |
| 132 |
class PlanarStrategy : public IPlanarStrategy<OutputStage> |
| 133 |
{ |
| 134 |
unsigned int m_kernel_rows, m_kernel_cols; |
| 135 |
unsigned int m_stride_rows, m_stride_cols; |
| 136 |
unsigned int m_output_rows; |
| 137 |
arm_gemm::VLType m_vl_type; |
| 138 |
|
| 139 |
protected: |
| 140 |
virtual bool get_kernel_packing_point(const unsigned int index, unsigned int &x, unsigned int &y) const |
| 141 |
{ |
| 142 |
// Get the kernel point to pack at the given index; return false to |
| 143 |
// indicate that this index (and all greater indices) is out of range. |
| 144 |
if (m_kernel_rows * m_kernel_cols <= index) |
| 145 |
return false; |
| 146 |
|
| 147 |
y = index % m_kernel_cols; |
| 148 |
x = index / m_kernel_cols; |
| 149 |
return true; |
| 150 |
} |
| 151 |
|
| 152 |
virtual interleaves::PackingArguments get_kernel_packing_arguments(void) const |
| 153 |
{ |
| 154 |
return interleaves::PackingArguments( |
| 155 |
m_kernel_rows, m_kernel_cols, sizeof(TWeight), |
| 156 |
false, sizeof(TAccum), true, // Don't pack the bias |
| 157 |
m_vl_type, sizeof(TAccum), 1, // Accumulator depth of 1 TODO |
| 158 |
[this] (unsigned int idx, unsigned int &x, unsigned int &y) -> bool |
| 159 |
{ return this->get_kernel_packing_point(idx, x, y); } |
| 160 |
); |
| 161 |
} |
| 162 |
|
| 163 |
public: |
| 164 |
PlanarStrategy( |
| 165 |
unsigned int kernel_rows, unsigned int kernel_cols, |
| 166 |
unsigned int stride_rows, unsigned int stride_cols, |
| 167 |
unsigned int output_rows, |
| 168 |
arm_gemm::VLType vl_type |
| 169 |
) : m_kernel_rows(kernel_rows), m_kernel_cols(kernel_cols), |
| 170 |
m_stride_rows(stride_rows), m_stride_cols(stride_cols), |
| 171 |
m_output_rows(output_rows), m_vl_type(vl_type) |
| 172 |
{ |
| 173 |
} |
| 174 |
|
| 175 |
unsigned int get_output_rows(void) const override { return m_output_rows; } |
| 176 |
arm_gemm::VLType get_vl_type(void) const override { return m_vl_type; } |
| 177 |
|
| 178 |
size_t get_storage_size(const DepthwiseArgs &args) const override |
| 179 |
{ |
| 180 |
return interleaves::get_storage_size_generic(this->get_kernel_packing_arguments(), args); |
| 181 |
} |
| 182 |
|
| 183 |
void pack_parameters( |
| 184 |
const DepthwiseArgs &args, void *buffer, |
| 185 |
const void *biases, const OutputStage &, |
| 186 |
const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 187 |
) const override |
| 188 |
{ |
| 189 |
interleaves::pack_parameters_generic( |
| 190 |
this->get_kernel_packing_arguments(), args, |
| 191 |
buffer, biases, weights, ld_weight_col, ld_weight_row |
| 192 |
); |
| 193 |
} |
| 194 |
|
| 195 |
using KernelType = typename PlanarKernelType<TInput, TWeight, TOutput, TAccum, OutputStage>::Type; |
| 196 |
virtual KernelType get_kernel(void) const = 0; |
| 197 |
}; |
| 198 |
|
| 199 |
|
| 200 |
namespace { |
| 201 |
|
| 202 |
template <typename T> |
| 203 |
struct OutputRowPtrsElement |
| 204 |
{ |
| 205 |
struct Workspace |
| 206 |
{ |
| 207 |
T **output_row_ptrs; |
| 208 |
size_t *output_ld_cols; |
| 209 |
size_t *output_ld_vls; // Stride between vectors of channels |
| 210 |
T *output_padding_buffer; |
| 211 |
}; |
| 212 |
|
| 213 |
// On some implementations there is a significant performance benefit to |
| 214 |
// aligning the padding buffer to a 1024 byte boundary. This routine |
| 215 |
// adds as much padding as needed to an arbitrary input pointer and |
| 216 |
// returns an aligned void *. |
| 217 |
static constexpr intptr_t BUFFER_ALIGNMENT=1024; |
| 218 |
|
| 219 |
template <typename ptr_T> |
| 220 |
static void *do_align(ptr_T in) |
| 221 |
{ |
| 222 |
intptr_t v = reinterpret_cast<intptr_t>(in); |
| 223 |
intptr_t odds = v & (BUFFER_ALIGNMENT-1); |
| 224 |
intptr_t pad = odds ? (BUFFER_ALIGNMENT - odds) : 0; |
| 225 |
|
| 226 |
return reinterpret_cast<void *>(v + pad); |
| 227 |
} |
| 228 |
|
| 229 |
template <typename OutputStage> |
| 230 |
static size_t get_element_size(const WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage> &args) |
| 231 |
{ |
| 232 |
// We need one pointer and stride for each row of output, and an |
| 233 |
// additional blob of memory into which padded stores can go. Allow |
| 234 |
// extra space so that this padding buffer can be aligned at both ends. |
| 235 |
return args.strategy->get_output_rows() * (sizeof(T *) + 2*sizeof(size_t)) + |
| 236 |
get_vector_length<char>(args.strategy->get_vl_type()) + BUFFER_ALIGNMENT*2; |
| 237 |
} |
| 238 |
|
| 239 |
template <typename WorkspaceType, typename OutputStage> |
| 240 |
static void *initialise(WorkspaceType *ws, void *buffer, |
| 241 |
const WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage> &args) |
| 242 |
{ |
| 243 |
const auto n_rows = args.strategy->get_output_rows(); |
| 244 |
ws->output_row_ptrs = reinterpret_cast<T **>(buffer); |
| 245 |
ws->output_ld_cols = reinterpret_cast<size_t *>(ws->output_row_ptrs + n_rows); |
| 246 |
ws->output_ld_vls = ws->output_ld_cols + n_rows; |
| 247 |
ws->output_padding_buffer = reinterpret_cast<T *>(do_align(ws->output_ld_vls + n_rows)); |
| 248 |
return do_align(ws->output_padding_buffer + get_vector_length<T>(args.strategy->get_vl_type())); |
| 249 |
} |
| 250 |
}; |
| 251 |
|
| 252 |
} // namespace {anonymous} |
| 253 |
|
| 254 |
|
| 255 |
template <typename TInput, typename TWeight=TInput, typename TOutput=TInput, |
| 256 |
typename TAccum=typename DefaultTAccum<TOutput>::Type, |
| 257 |
typename OutputStage=typename DefaultOutputStage<TOutput>::Type> |
| 258 |
class DepthwisePlanar : public DepthwiseCommon<TInput, TWeight, TOutput> |
| 259 |
{ |
| 260 |
using Parent = DepthwiseCommon<TInput, TWeight, TOutput>; |
| 261 |
using StrategyType = IPlanarStrategy<OutputStage>; |
| 262 |
using WorkspaceManager = Workspace< |
| 263 |
OutputRowPtrsElement<TOutput>, |
| 264 |
ActivationsElement<TAccum, OutputStage> |
| 265 |
>; |
| 266 |
using WorkspaceType = typename WorkspaceManager::WorkspaceType; |
| 267 |
|
| 268 |
std::unique_ptr<StrategyType> m_strat; |
| 269 |
const TAccum *m_bias; |
| 270 |
OutputStage m_os; |
| 271 |
|
| 272 |
public: |
| 273 |
DepthwisePlanar(StrategyType *const strat, const DepthwiseArgs &args, const OutputStage &os = {}) |
| 274 |
: Parent(args), m_strat(strat), m_bias(nullptr), m_os(os) |
| 275 |
{ |
| 276 |
} |
| 277 |
|
| 278 |
DepthwisePlanar(DepthwisePlanar &) = delete; |
| 279 |
DepthwisePlanar &operator=(DepthwisePlanar &) = delete; |
| 280 |
|
| 281 |
size_t get_storage_size(void) const override |
| 282 |
{ |
| 283 |
return m_strat->get_storage_size(this->m_args); |
| 284 |
} |
| 285 |
|
| 286 |
void pack_parameters( |
| 287 |
void *buffer, const void *biases, |
| 288 |
const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 289 |
) override |
| 290 |
{ |
| 291 |
m_strat->pack_parameters(this->m_args, buffer, biases, {}, weights, ld_weight_col, ld_weight_row); |
| 292 |
this->m_bias = reinterpret_cast<const TAccum *>(biases); |
| 293 |
depthwise_depthfirst::stash_bias(this->m_os, biases); |
| 294 |
} |
| 295 |
|
| 296 |
size_t get_working_size(unsigned int n_threads) const override |
| 297 |
{ |
| 298 |
return this->get_working_size_per_thread() * n_threads; |
| 299 |
} |
| 300 |
|
| 301 |
protected: |
| 302 |
/* Compute the amount of working space required for a single thread. */ |
| 303 |
virtual size_t get_working_size_per_thread(void) const |
| 304 |
{ |
| 305 |
return WorkspaceManager::get_sizeof_workspace( |
| 306 |
WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage>(m_strat.get(), this->m_args, m_os)); |
| 307 |
} |
| 308 |
|
| 309 |
/* Initialise the working space for a thread. */ |
| 310 |
virtual void initialise_working_space(void *buffer) const |
| 311 |
{ |
| 312 |
WorkspaceManager::initialise( |
| 313 |
buffer, |
| 314 |
WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage>(m_strat.get(), this->m_args, m_os) |
| 315 |
); |
| 316 |
} |
| 317 |
|
| 318 |
/* Execute the kernel for a given chunk of work. */ |
| 319 |
virtual void execute_kernel( |
| 320 |
const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 321 |
unsigned int pad_top, unsigned int valid_input_rows, |
| 322 |
unsigned int pad_left, unsigned int valid_input_cols, |
| 323 |
const TWeight *weights, const TAccum *bias, |
| 324 |
TOutput *outptr, size_t ld_out_row, size_t ld_out_col, size_t ld_out_vl, |
| 325 |
unsigned int valid_output_rows, unsigned int valid_output_cols, |
| 326 |
unsigned int first_channel, unsigned int valid_channels, |
| 327 |
WorkspaceType *ws |
| 328 |
) const |
| 329 |
{ |
| 330 |
// Initialise the output pointers |
| 331 |
for (auto i = 0u; i < m_strat->get_output_rows(); i++) |
| 332 |
{ |
| 333 |
// Point at the output tensor for all valid rows; otherwise point at the |
| 334 |
// padding buffer. |
| 335 |
ws->output_row_ptrs[i] = i < valid_output_rows ? outptr : ws->output_padding_buffer; |
| 336 |
ws->output_ld_cols[i] = i < valid_output_rows ? ld_out_col : 0; |
| 337 |
ws->output_ld_vls[i] = i < valid_output_rows ? ld_out_vl : 0; |
| 338 |
outptr += ld_out_row; |
| 339 |
} |
| 340 |
|
| 341 |
// Execute the kernel |
| 342 |
PlanarKernelType<TInput, TWeight, TOutput, TAccum, OutputStage>::template execute<WorkspaceType>( |
| 343 |
reinterpret_cast<const PlanarStrategy<TInput, TWeight, TOutput, TAccum, OutputStage> *>(m_strat.get())->get_kernel(), |
| 344 |
inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 345 |
pad_top, valid_input_rows, pad_left, valid_input_cols, |
| 346 |
weights, bias, |
| 347 |
ws->output_row_ptrs, ws->output_ld_cols, ws->output_ld_vls, |
| 348 |
valid_output_cols, first_channel, valid_channels, |
| 349 |
this->m_os, ws |
| 350 |
); |
| 351 |
} |
| 352 |
|
| 353 |
void execute_internal( |
| 354 |
const DepthwiseArgs &args, |
| 355 |
const void *input, |
| 356 |
size_t ld_input_col, |
| 357 |
size_t ld_input_row, |
| 358 |
size_t ld_input_batch, |
| 359 |
const void *parameters, |
| 360 |
void *output, |
| 361 |
size_t ld_output_col, |
| 362 |
size_t ld_output_row, |
| 363 |
size_t ld_output_batch, |
| 364 |
void *working_space, |
| 365 |
unsigned int thread_id, |
| 366 |
unsigned int n_threads |
| 367 |
) const override |
| 368 |
{ |
| 369 |
// Get and initialise the working space for this thread. |
| 370 |
void *thread_working_space = |
| 371 |
static_cast<uint8_t *>(working_space) + thread_id * this->get_working_size_per_thread(); |
| 372 |
this->initialise_working_space(thread_working_space); |
| 373 |
auto ws = reinterpret_cast<WorkspaceType *>(thread_working_space); |
| 374 |
|
| 375 |
const auto n_output_channels = args.input_channels * args.channel_multiplier; |
| 376 |
const auto vl = get_vector_length<TAccum>(m_strat->get_vl_type()); |
| 377 |
|
| 378 |
// Get typed pointers |
| 379 |
auto input_batch = reinterpret_cast<const TInput *>(input); |
| 380 |
auto output_batch = reinterpret_cast<TOutput *>(output); |
| 381 |
auto weights = reinterpret_cast<const TWeight *>(parameters); |
| 382 |
|
| 383 |
// Iterate over batches |
| 384 |
for (auto batches = args.n_batches; batches; batches--) |
| 385 |
{ |
| 386 |
// NOTE: Other loop orderings are possible and it would be worth |
| 387 |
// investigating them. |
| 388 |
|
| 389 |
// Within a batch, stripe threads across rows. |
| 390 |
for (auto start_output_i = thread_id * m_strat->get_output_rows(); |
| 391 |
start_output_i < args.output_rows; |
| 392 |
start_output_i += n_threads * m_strat->get_output_rows()) |
| 393 |
{ |
| 394 |
// Determine what (if any padding) is required on the top/bottom of |
| 395 |
// this row of the convolution. |
| 396 |
const int start_input_i = start_output_i * args.stride_rows - args.padding.top; |
| 397 |
const unsigned int input_pad_top = start_input_i < 0 ? -start_input_i : 0; |
| 398 |
const unsigned int input_i = start_input_i < 0 ? 0 : start_input_i; |
| 399 |
const unsigned int valid_input_rows = input_i > args.input_rows ? 0 : args.input_rows - input_i; |
| 400 |
const unsigned int valid_output_rows = args.output_rows - start_output_i; |
| 401 |
|
| 402 |
auto inptr_row = input_batch + input_i*ld_input_row; |
| 403 |
auto outptr_row = output_batch + start_output_i * ld_output_row; |
| 404 |
|
| 405 |
// Execute the kernel |
| 406 |
this->execute_kernel( |
| 407 |
inptr_row, ld_input_row, ld_input_col, vl, |
| 408 |
input_pad_top, valid_input_rows, args.padding.left, args.input_cols, |
| 409 |
weights, this->m_bias, |
| 410 |
outptr_row, ld_output_row, ld_output_col, vl, |
| 411 |
valid_output_rows, args.output_cols, |
| 412 |
0 /* first channel */, n_output_channels, |
| 413 |
ws |
| 414 |
); |
| 415 |
} |
| 416 |
|
| 417 |
// Update the input and output pointers to account for batch |
| 418 |
input_batch += ld_input_batch; |
| 419 |
output_batch += ld_output_batch; |
| 420 |
} |
| 421 |
} |
| 422 |
}; |
| 423 |
|
| 424 |
} // namespace depthwise |
| 425 |
} // namespace arm_conv |