Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2020 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 | #pragma once |
| 25 | |
| 26 | #include "arm_gemm.hpp" |
| 27 | #include "utils.hpp" |
| 28 | |
| 29 | #include "mergeresults.hpp" |
| 30 | #include "transform.hpp" |
| 31 | |
| 32 | #ifdef CYCLE_PROFILING |
| 33 | #include "profiler.hpp" |
| 34 | #endif |
| 35 | |
| 36 | #include <algorithm> |
| 37 | #include <cassert> |
| 38 | |
| 39 | // Some macros used to decide how much working space to allocate. |
| 40 | // Round allocations up to the next cache line. |
| 41 | #define ALLOC_ROUND 64 |
| 42 | #define ROUND_UP(x) ((((x) + ALLOC_ROUND-1) / ALLOC_ROUND) * ALLOC_ROUND) |
| 43 | |
| 44 | // Implementation of the GemmCommon abstract class. |
| 45 | // |
| 46 | // This implementation interleaves the source matrices in blocks - good for |
| 47 | // larger matrices. |
| 48 | namespace arm_gemm { |
| 49 | |
| 50 | template<typename strategy, typename To, typename Tr> |
| 51 | class GemmInterleaved2d : public GemmCommon<To, Tr> { |
| 52 | typedef typename strategy::operand_type Toi; |
| 53 | typedef typename strategy::result_type Tri; |
| 54 | |
| 55 | /* const properties set by constructor */ |
| 56 | const CPUInfo * const _ci; |
| 57 | |
| 58 | const unsigned int _Msize; |
| 59 | const unsigned int _Nsize; |
| 60 | const unsigned int _Ksize; |
| 61 | |
| 62 | const unsigned int _nbatches; |
| 63 | const unsigned int _nmulti; |
| 64 | |
| 65 | const bool _trA; |
| 66 | const bool _trB; |
| 67 | |
| 68 | const Activation _act; |
| 69 | |
| 70 | const int _maxthreads; |
| 71 | int _nthreads; |
| 72 | |
| 73 | /* Blocking info */ |
| 74 | unsigned int _k_block=0; |
| 75 | unsigned int _x_block=0; |
| 76 | |
| 77 | unsigned int _Mround_div=0; |
| 78 | unsigned int _Mround=0; |
| 79 | unsigned int _Nround_div=0; |
| 80 | unsigned int _Nround=0; |
| 81 | |
| 82 | /* Working space, pretransposed buffer */ |
| 83 | void *_working_space=nullptr; |
| 84 | |
| 85 | /* We will need to walk through the blocks of B in a few contexts, so |
| 86 | * factor that out. */ |
| 87 | class blockwalker { |
| 88 | private: |
| 89 | /* Size loops, etc. based on our parent's configuration */ |
| 90 | const GemmInterleaved2d<strategy, To, Tr> &_parent; |
| 91 | |
| 92 | /* K, X and multi parameters for current iteration. */ |
| 93 | unsigned int _k0=0, _x0=0, _xmin=0, _xmax=0, _multi=0; |
| 94 | |
| 95 | unsigned int _index=0; |
| 96 | bool _done=false; |
| 97 | bool _newkblock=true; |
| 98 | bool _newmulti=true; |
| 99 | |
| 100 | public: |
| 101 | blockwalker(const GemmInterleaved2d<strategy, To, Tr> &parent) |
| 102 | : _parent(parent) |
| 103 | , _xmax { parent._Nsize } |
| 104 | { } |
| 105 | |
| 106 | blockwalker(const GemmInterleaved2d<strategy, To, Tr> &parent, unsigned int x0, unsigned int xmax) |
| 107 | : _parent(parent) |
| 108 | , _x0 { x0 } |
| 109 | , _xmin { x0 } |
| 110 | , _xmax { xmax } |
| 111 | { |
| 112 | assert(_x0 <= _xmax); |
| 113 | } |
| 114 | |
| 115 | unsigned int xmax() { |
| 116 | return std::min(_x0 + _parent._x_block, _xmax); |
| 117 | } |
| 118 | |
| 119 | unsigned int kmax() { |
| 120 | return std::min(_k0 + _parent._k_block, _parent._Ksize); |
| 121 | } |
| 122 | |
| 123 | /* Advance to the next block, return false at the end. */ |
| 124 | bool advance(void) { |
| 125 | if (_done) { |
| 126 | return false; |
| 127 | } |
| 128 | |
| 129 | _newkblock=false; |
| 130 | _x0 += _parent._x_block; |
| 131 | if (_x0 >= _xmax) { |
| 132 | _x0=_xmin; |
| 133 | _k0 += _parent._k_block; |
| 134 | if (_k0 >= _parent._Ksize) { |
| 135 | _k0=0; |
| 136 | _multi++; |
| 137 | if (_multi >= _parent._nmulti) { |
| 138 | _done=true; |
| 139 | return false; |
| 140 | } |
| 141 | _newmulti=true; |
| 142 | } |
| 143 | _newkblock=true; |
| 144 | } |
| 145 | _index++; |
| 146 | |
| 147 | return true; |
| 148 | } |
| 149 | |
| 150 | unsigned int k0(void) { return _k0; } |
| 151 | unsigned int x0(void) { return _x0; } |
| 152 | unsigned int multi(void) { return _multi; } |
| 153 | unsigned int index(void) { return _index; } |
| 154 | bool done(void) { return _done; } |
| 155 | bool newkblock(void) { return _newkblock; } |
| 156 | }; |
| 157 | |
| 158 | // A working size: One of these needed, regardless of thread count. Divided according to window. |
| 159 | size_t get_a_working_size() const { |
| 160 | return ROUND_UP(sizeof(Toi) * _k_block * _Mround * _nbatches) * 2; |
| 161 | } |
| 162 | |
| 163 | // B working size: 0, 1 or 3 of these needed depending on pretransposed and threading settings. |
| 164 | size_t get_b_working_size() const { |
| 165 | return ROUND_UP(sizeof(Toi) * _x_block * _k_block); |
| 166 | } |
| 167 | |
| 168 | // C working size: One needed per thread. |
| 169 | size_t get_c_working_size() const { |
| 170 | return ROUND_UP(sizeof(Tri) * _x_block * strategy::out_height()); |
| 171 | } |
| 172 | |
Georgios Pinitas | 5aa1a0b | 2020-07-02 20:02:20 +0100 | [diff] [blame] | 173 | void execute_transpose(unsigned int m_start, unsigned int m_end, unsigned int n_start, unsigned int n_end, int threadid, int, int nthreadid) { |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 174 | strategy strat(_ci); |
| 175 | |
| 176 | /* Translate 'start' and 'end' into a position within the batches and rows. */ |
| 177 | const unsigned int window_per_batch = _Mround / strategy::out_height(); |
| 178 | unsigned int batch_0 = m_start / window_per_batch; |
| 179 | unsigned int batch_end = m_end / window_per_batch; |
| 180 | |
| 181 | /* Compute the M values to operate on */ |
| 182 | unsigned int m_0 = (m_start - (batch_0 * window_per_batch)) * strategy::out_height(); |
| 183 | unsigned int m_max = (m_end - (batch_end * window_per_batch)) * strategy::out_height(); |
| 184 | |
| 185 | unsigned int n_0 = std::min(this->_Nsize, strategy::out_width() * n_start); |
| 186 | unsigned int n_max = std::min(this->_Nsize, strategy::out_width() * n_end); |
| 187 | |
| 188 | blockwalker current(*this, n_0, n_max); |
| 189 | |
| 190 | /* get workspace as int8_t */ |
| 191 | assert(_working_space); |
| 192 | int8_t *working_space_bytes = reinterpret_cast<int8_t *>(_working_space); |
| 193 | |
| 194 | auto c_panel_start = working_space_bytes; |
| 195 | auto a_panel_start = c_panel_start + get_c_working_size() * _maxthreads; |
| 196 | auto b_panel_start = a_panel_start + get_a_working_size() * _maxthreads; |
| 197 | |
| 198 | auto c_panel = reinterpret_cast<Tri *>(c_panel_start + get_c_working_size() * threadid); |
| 199 | auto a_panel = reinterpret_cast<Toi *>(a_panel_start + get_a_working_size() * nthreadid); |
| 200 | auto b_panel = reinterpret_cast<Toi *>(b_panel_start + get_b_working_size() * threadid); |
| 201 | |
| 202 | |
| 203 | // newkblock() is always true on the first iteration, so this will be set properly on the first loop. |
| 204 | |
| 205 | int kern_k = 0; |
| 206 | for (;!current.done();current.advance()) { |
| 207 | const int bblocks = iceildiv(current.xmax() - current.x0(), strategy::out_width()); |
| 208 | /* |
| 209 | * The entirity of A^kblock is transpose upfront and computed against individual |
| 210 | * blocks of B (xblock) |
| 211 | * |
| 212 | * Therefore, we only need to retranspose when k_block progresses |
| 213 | */ |
| 214 | if (current.newkblock()) { |
| 215 | for (unsigned int batch = batch_0; batch <= batch_end; batch++) { |
| 216 | unsigned int first_m = (batch == batch_0) ? m_0 : 0; |
| 217 | unsigned int last_m = (batch == batch_end) ? m_max : _Msize; |
| 218 | |
| 219 | if (first_m >= last_m) |
| 220 | continue; |
| 221 | |
| 222 | auto a_thread_panel_in = this->_Aptr |
| 223 | + (batch * this->_A_batch_stride) |
| 224 | + (current.multi() * this->_A_multi_stride); |
| 225 | |
| 226 | auto a_thread_panel_out = a_panel + ((batch * _Mround + first_m) * _k_block); |
| 227 | |
| 228 | strat.transforms.PrepareA( |
| 229 | a_thread_panel_out, |
| 230 | a_thread_panel_in, |
| 231 | this->_lda, |
| 232 | first_m, |
| 233 | last_m, |
| 234 | current.k0(), |
| 235 | current.kmax(), |
| 236 | _trA); |
| 237 | } |
| 238 | |
| 239 | kern_k = iceildiv(current.kmax() - current.k0(), strategy::k_unroll()); |
| 240 | kern_k *= strat.k_unroll(); |
| 241 | } |
| 242 | |
| 243 | auto *b_panel_in = this->_Bptr + (current.multi() * this->_B_multi_stride); |
| 244 | |
| 245 | strat.transforms.PrepareB( |
| 246 | b_panel, //dst |
| 247 | b_panel_in, //src |
| 248 | this->_ldb, |
| 249 | current.x0(), //idx from |
| 250 | current.xmax(), //idx to |
| 251 | current.k0(), |
| 252 | current.kmax(), |
| 253 | _trB); |
| 254 | |
| 255 | //Iterate over the batches |
| 256 | for (unsigned int batch = batch_0; batch <= batch_end; batch++) { |
| 257 | unsigned int first_m = (batch == batch_0) ? m_0 : 0; |
| 258 | unsigned int last_m = (batch == batch_end) ? m_max : _Msize; |
| 259 | |
| 260 | if (first_m >= last_m) |
| 261 | continue; |
| 262 | |
| 263 | const Toi *a_ptr = a_panel + (batch * _Mround + first_m) * _k_block; |
| 264 | |
| 265 | |
| 266 | //Iterate over the inerleaved rows of the packed A matrix |
| 267 | for (unsigned int y=first_m; y<last_m; y+=strategy::out_height()) { |
| 268 | unsigned int ymax = std::min(_Msize, y + strategy::out_height()); |
| 269 | |
| 270 | strat.kernel(a_ptr, b_panel, c_panel, 1, bblocks, kern_k); |
| 271 | a_ptr += (strategy::out_height() * kern_k); |
| 272 | |
| 273 | const bool first_pass = current.k0()==0; |
| 274 | const bool last_pass = current.kmax()==_Ksize; |
| 275 | |
| 276 | auto c_panel_out = this->_Cptr |
| 277 | + this->_C_batch_stride * batch |
| 278 | + this->_C_multi_stride * current.multi(); |
| 279 | |
| 280 | auto bias = (first_pass && this->_bias) |
| 281 | ? this->_bias + (current.multi() * this->_bias_multi_stride) |
| 282 | : nullptr; |
| 283 | |
| 284 | auto act = last_pass ? _act : Activation(); |
| 285 | |
| 286 | strat.transforms.Merge( |
| 287 | c_panel_out, |
| 288 | c_panel, |
| 289 | this->_ldc, |
| 290 | y, |
| 291 | ymax, |
| 292 | current.x0(), |
| 293 | current.xmax(), |
| 294 | bias, |
| 295 | act, |
| 296 | !first_pass); //Append |
| 297 | } |
| 298 | } |
| 299 | } |
| 300 | } |
| 301 | public: |
| 302 | GemmInterleaved2d(GemmInterleaved2d &) = delete; |
| 303 | GemmInterleaved2d & operator= (GemmInterleaved2d &) = delete; |
| 304 | |
| 305 | /* Constructor */ |
| 306 | /* Constructor */ |
| 307 | GemmInterleaved2d(const GemmArgs &args) |
| 308 | : _ci(args._ci) |
| 309 | , _Msize(args._Msize) |
| 310 | , _Nsize(args._Nsize) |
| 311 | , _Ksize(args._Ksize) |
| 312 | , _nbatches(args._nbatches) |
| 313 | , _nmulti(args._nmulti) |
| 314 | , _trA(args._trA) |
| 315 | , _trB(args._trB) |
| 316 | , _act(args._act) |
| 317 | , _maxthreads(args._maxthreads) |
| 318 | , _nthreads(args._maxthreads) |
| 319 | |
| 320 | // Work out the rounded size of M - needed for some buffers. |
| 321 | , _Mround_div ( iceildiv(_Msize, strategy::out_height()) ) |
| 322 | , _Mround ( _Mround_div * strategy::out_height() ) |
| 323 | |
| 324 | , _Nround_div ( iceildiv(_Nsize, strategy::out_width()) ) |
| 325 | , _Nround ( _Nround_div * strategy::out_width() ) |
| 326 | { |
| 327 | const unsigned int L1_size = _ci->get_L1_cache_size(); |
| 328 | const unsigned int L2_size = _ci->get_L2_cache_size(); |
| 329 | |
| 330 | assert(_maxthreads > 0); |
| 331 | |
| 332 | // Work out blocking parameters, or override from provided GemmConfig |
| 333 | if (args._cfg && args._cfg->inner_block_size) { |
| 334 | _k_block = args._cfg->inner_block_size; |
| 335 | } else { |
| 336 | // k_block: Find out how much of the larger array can be loaded into half the cache. |
| 337 | // This should account for associative caches. |
| 338 | _k_block = (L1_size / 2) / (sizeof(Toi) * (std::max(strategy::out_width(), strategy::out_height()))); |
| 339 | |
| 340 | // Needs to be (at least a single) multiple of the K unroll level. |
| 341 | _k_block /= strategy::k_unroll(); |
| 342 | _k_block = std::max(_k_block, 1U) * strategy::k_unroll(); |
| 343 | |
| 344 | // Now tune to presented problem size; this is how many blocks we need. |
| 345 | unsigned int num_k_blocks = iceildiv(_Ksize, _k_block); |
| 346 | |
| 347 | // So divide the space equally into that many blocks. |
| 348 | _k_block = iceildiv(_Ksize, num_k_blocks); |
| 349 | |
| 350 | // And round UP to the K unroll level required. |
| 351 | _k_block = iceildiv(_k_block, strategy::k_unroll()); |
| 352 | _k_block *= strategy::k_unroll(); |
| 353 | } |
| 354 | |
| 355 | if (args._cfg && args._cfg->outer_block_size) { |
| 356 | _x_block = args._cfg->outer_block_size; |
| 357 | } else { |
| 358 | // x_block: Work out how many rows (of length k_block) will fit in the L2 |
| 359 | // Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents. |
| 360 | _x_block = (((L2_size * 9) / 10) - (_k_block * sizeof(Toi) * (strategy::out_width() + strategy::out_height()))) / |
| 361 | (sizeof(Toi) * _k_block); |
| 362 | |
| 363 | // Needs to be (at least a single) multiple of the kernel output width. |
| 364 | _x_block /= strategy::out_width(); |
| 365 | _x_block = std::max(_x_block, 1U) * strategy::out_width(); |
| 366 | |
| 367 | // And tune to the presented problem size. |
| 368 | unsigned int num_x_blocks = iceildiv(_Nsize, _x_block); |
| 369 | _x_block = iceildiv(_Nsize, num_x_blocks); |
| 370 | |
| 371 | _x_block = iceildiv(_x_block, strategy::out_width()); |
| 372 | _x_block *= strategy::out_width(); |
| 373 | } |
| 374 | |
| 375 | // Work out the rounded size of M - needed for some buffers. |
| 376 | } |
| 377 | |
| 378 | // Interface implementation - Compulsory functions |
| 379 | ndrange_t get_window_size() const override { |
| 380 | unsigned m = (_Mround / strategy::out_height()) * _nbatches; |
| 381 | unsigned n = _Nround_div; |
| 382 | |
Georgios Pinitas | 5aa1a0b | 2020-07-02 20:02:20 +0100 | [diff] [blame] | 383 | return { m, n }; |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 384 | } |
| 385 | |
| 386 | // set_nthreads: pass on to buffer manager to avoid it waiting for non-existant threads. |
| 387 | void set_nthreads(int nthreads) override { |
| 388 | _nthreads = std::min(nthreads, _maxthreads); |
| 389 | } |
| 390 | |
| 391 | void execute(const ndcoord_t& work_range, const ndcoord_t& thread_locator, int threadid) override { |
| 392 | /* |
| 393 | * This particular GEMM implementation can only be broken up over the M & N |
| 394 | * dimensions, we inform the frame work of this limitation via the get_window_size function |
| 395 | */ |
Joseph Dobson | 6f8b17d | 2020-02-11 19:32:11 +0000 | [diff] [blame] | 396 | const auto m_start = work_range.get_position(0); |
| 397 | const auto n_start = work_range.get_position(1); |
| 398 | const auto m_size = work_range.get_size(0); |
| 399 | const auto n_size = work_range.get_size(1); |
| 400 | const auto m_end = m_start + m_size; |
| 401 | const auto n_end = n_start + n_size; |
| 402 | |
| 403 | const auto m_threadid = thread_locator.get_position(0); |
| 404 | const auto n_threadid = thread_locator.get_position(1); |
| 405 | |
| 406 | execute_transpose(m_start, m_end, n_start, n_end, threadid, m_threadid, n_threadid); |
| 407 | } |
| 408 | |
| 409 | std::size_t get_working_size()const override { |
| 410 | /* |
| 411 | * Because we do not know how schedular will break up |
| 412 | * the task, we need to ensure that alloc enough |
| 413 | * space to be able to handle the case where every thread |
| 414 | * is parallelised across B AND also every thrread is parallelised across A |
| 415 | * |
| 416 | * If we parallelise across A, then we only need one buffer of A and 64 buffers of B |
| 417 | * If we parallelise across B, then we only need 64 buffer of B and |
| 418 | */ |
| 419 | return get_c_working_size() * _maxthreads |
| 420 | + get_a_working_size() * _maxthreads |
| 421 | + get_b_working_size() * _maxthreads |
| 422 | + 64; //to account for cacheline alignment |
| 423 | } |
| 424 | |
| 425 | |
| 426 | void set_working_space(void *working_space) override { |
| 427 | // Make sure everything ends up cache line aligned |
| 428 | int8_t *working_space_bytes = reinterpret_cast<int8_t *>(working_space); |
| 429 | intptr_t working_space_int = reinterpret_cast<intptr_t>(working_space); |
| 430 | |
| 431 | size_t diff=0; |
| 432 | |
| 433 | if (working_space_int & 0x3F) { |
| 434 | diff = 0x40 - (working_space_int & 0x3F); |
| 435 | } |
| 436 | |
| 437 | working_space_bytes += diff; |
| 438 | |
| 439 | _working_space = reinterpret_cast<void *>(working_space_bytes); |
| 440 | } |
| 441 | |
| 442 | ~GemmInterleaved2d() override { } |
| 443 | }; |
| 444 | |
| 445 | } // namespace arm_gemm |