Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 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 <assert.h> |
| 27 | |
| 28 | #include <algorithm> |
| 29 | |
| 30 | #include "arm_gemm.hpp" |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 31 | #include "bias_adder.hpp" |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 32 | #include "ndrange.hpp" |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 33 | #include "utils.hpp" |
| 34 | |
| 35 | #include "mergeresults.hpp" |
| 36 | #include "transform.hpp" |
| 37 | |
| 38 | #ifdef CYCLE_PROFILING |
| 39 | #include "profiler.hpp" |
| 40 | #endif |
| 41 | |
| 42 | namespace arm_gemm { |
| 43 | |
| 44 | // Implementation of the GemmCommon abstract class. |
| 45 | template<typename strategy, typename To, typename Tr> |
| 46 | class GemmHybrid : public GemmCommon<To, Tr> { |
| 47 | typedef typename strategy::operand_type Toi; |
| 48 | typedef typename strategy::result_type Tri; |
| 49 | |
| 50 | /* const properties set by constructor */ |
| 51 | const CPUInfo * const _ci; |
| 52 | |
| 53 | const unsigned int _Msize; |
| 54 | const unsigned int _Nsize; |
| 55 | const unsigned int _Ksize; |
| 56 | |
| 57 | const unsigned int _nbatches; |
| 58 | const unsigned int _nmulti; |
| 59 | |
| 60 | const bool _trB; |
| 61 | |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 62 | const Activation _act; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 63 | |
| 64 | /* Blocking info */ |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 65 | const unsigned int _k_block; |
| 66 | const unsigned int _n_block; |
| 67 | const unsigned int _Mround; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 68 | |
| 69 | /* Pretransposed buffer. */ |
| 70 | const Toi *_B_transposed=nullptr; |
| 71 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 72 | const NDRange<4> _window_range; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 73 | |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 74 | static unsigned int compute_k_block(const GemmArgs &args) { |
| 75 | // Some kernels don't support append mode - these can't do K blocking at all. |
| 76 | if (!strategy::supports_append()) { |
| 77 | return args._Ksize; |
| 78 | } |
| 79 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 80 | if (args._cfg && args._cfg->inner_block_size) { |
| 81 | return args._cfg->inner_block_size; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 82 | } |
| 83 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 84 | const unsigned int L1_size = args._ci->get_L1_cache_size(); |
| 85 | |
| 86 | // k_block: Find out how much of the larger array can be loaded into half the cache. |
| 87 | // This should account for associative caches. |
| 88 | unsigned int k_block = (L1_size / 2) / (sizeof(Toi) * (std::max(strategy::out_width(), strategy::out_height()))); |
| 89 | |
| 90 | // Needs to be (at least a single) multiple of the K unroll level. |
| 91 | k_block /= strategy::k_unroll(); |
| 92 | k_block = std::max(k_block, 1U) * strategy::k_unroll(); |
| 93 | |
| 94 | // Now tune to presented problem size; this is how many blocks we need. |
| 95 | unsigned int numk_blocks = iceildiv(args._Ksize, k_block); |
| 96 | |
| 97 | // So divide the space equally into that many blocks. |
| 98 | k_block = iceildiv(args._Ksize, numk_blocks); |
| 99 | |
| 100 | // And round UP to the K unroll level required. |
| 101 | k_block = roundup(k_block, strategy::k_unroll()); |
| 102 | |
| 103 | return k_block; |
| 104 | } |
| 105 | |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 106 | static unsigned int compute_n_block(const GemmArgs &args) { |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 107 | if (args._cfg && args._cfg->outer_block_size) { |
| 108 | return args._cfg->outer_block_size; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 109 | } |
| 110 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 111 | const unsigned int k_block = compute_k_block(args); |
| 112 | const unsigned int L2_size = args._ci->get_L2_cache_size(); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 113 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 114 | // n_block: Work out how many rows (of length k_block) will fit in the L2 |
| 115 | // Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents. |
| 116 | unsigned int n_block = (((L2_size * 9) / 10) - (k_block * sizeof(Toi) * (strategy::out_width() + strategy::out_height()))) / |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 117 | (sizeof(Toi) * k_block); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 118 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 119 | // Needs to be (at least a single) multiple of the kernel output width. |
| 120 | n_block /= strategy::out_width(); |
| 121 | n_block = std::max(n_block, 1U) * strategy::out_width(); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 122 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 123 | // And tune to the presented problem size. |
| 124 | unsigned int numblocks = iceildiv(args._Nsize, n_block); |
| 125 | n_block = iceildiv(args._Nsize, numblocks); |
| 126 | n_block = roundup(n_block, strategy::out_width()); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 127 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 128 | return n_block; |
| 129 | } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 130 | |
| 131 | public: |
| 132 | GemmHybrid(GemmHybrid &) = delete; |
| 133 | GemmHybrid & operator= (GemmHybrid &) = delete; |
| 134 | |
| 135 | /* Constructor */ |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 136 | GemmHybrid(const GemmArgs &args) |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 137 | : _ci(args._ci), _Msize(args._Msize), _Nsize(args._Nsize), _Ksize(args._Ksize), |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 138 | _nbatches(args._nbatches), _nmulti(args._nmulti), _trB(args._trB), |
| 139 | _act(args._act), |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 140 | _k_block(compute_k_block(args)), _n_block(compute_n_block(args)), |
| 141 | _Mround(roundup(args._Msize, strategy::out_height())), |
| 142 | _window_range(iceildiv(args._Msize, strategy::out_height()), _nbatches, iceildiv(_Nsize, _n_block), _nmulti) { } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 143 | |
| 144 | // Interface implementation - Compulsory functions |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 145 | unsigned int get_window_size() const override { |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 146 | return _window_range.total_size(); |
| 147 | } |
| 148 | |
| 149 | // This kernel can always be dynamically scheduled. |
| 150 | bool supports_dynamic_scheduling() const override { |
| 151 | return true; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 152 | } |
| 153 | |
| 154 | // Execute |
| 155 | void execute(unsigned int start, unsigned int end, int threadid) override { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 156 | UNUSED(threadid); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 157 | #ifdef CYCLE_PROFILING |
| 158 | profiler prof; |
| 159 | #endif |
| 160 | strategy strat(_ci); |
| 161 | |
| 162 | /* Make sure we've been set up correctly. */ |
| 163 | assert(_B_transposed); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 164 | static_assert(std::is_same<To, Toi>::value, "gemm_native: Operand types must be the same."); |
| 165 | static_assert(std::is_same<Tr, Tri>::value, "gemm_native: Result types must be the same."); |
| 166 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 167 | /* For now, each work item implies all the K for a given output |
| 168 | * pixel (so we don't need to synchronize access to the output |
| 169 | * array). So separate the loop over K blocks here. */ |
| 170 | for (unsigned int k0=0; k0<_Ksize; k0+=_k_block) { |
| 171 | unsigned int kmax = std::min(k0 + _k_block, _Ksize); |
| 172 | unsigned int kern_k = roundup(kmax-k0, strategy::k_unroll()); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 173 | |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 174 | const bool first_pass = (k0 == 0); |
| 175 | const bool last_pass = (kmax == _Ksize); |
| 176 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 177 | auto p = _window_range.iterator(start, end); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 178 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 179 | if (p.done()) { |
| 180 | return; |
| 181 | } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 182 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 183 | do { |
| 184 | const unsigned int m_start = p.dim(0) * strategy::out_height(); |
| 185 | const unsigned int m_end = std::min(p.dim0_max() * strategy::out_height(), _Msize); |
| 186 | const unsigned int batch = p.dim(1); |
| 187 | const unsigned int n0 = p.dim(2) * _n_block; |
| 188 | const unsigned int nmax = std::min(n0 + _n_block, _Nsize); |
| 189 | const unsigned int multi = p.dim(3); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 190 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 191 | const Toi *b_panel = _B_transposed + |
| 192 | (multi * roundup(_Nsize, strategy::out_width()) * roundup(_Ksize, strategy::k_unroll())) + |
| 193 | (k0 * roundup(_Nsize, strategy::out_width())) + |
| 194 | (n0 * kern_k); |
| 195 | |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 196 | #ifdef CYCLE_PROFILING |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 197 | auto p = prof.ScopedProfiler(PROFILE_KERNEL, (m_end - m_start) * kern_k * roundup(nmax-n0, strategy::out_width())); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 198 | #endif |
| 199 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 200 | strat.kernel(this->_Aptr + (multi * this->_A_multi_stride) + (batch * this->_A_batch_stride) + (m_start * this->_lda) + k0, this->_lda, |
| 201 | b_panel, |
| 202 | this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (m_start * this->_ldc) + n0, this->_ldc, |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 203 | (m_end - m_start), (nmax - n0), kmax-k0, |
| 204 | (strategy::supports_bias() && first_pass && this->_bias) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr, |
| 205 | last_pass ? _act : Activation(), !first_pass); |
| 206 | |
| 207 | // Add bias externally if needed |
| 208 | if (!strategy::supports_bias() && this->_bias && first_pass) { |
| 209 | bias_adder(this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (m_start * this->_ldc) + n0, this->_ldc, |
| 210 | this->_bias + (multi * this->_bias_multi_stride) + n0, |
| 211 | (m_end - m_start), (nmax - n0)); |
| 212 | } |
| 213 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 214 | } while (p.next_dim1()); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 215 | } |
| 216 | } |
| 217 | |
| 218 | // Interface implementation - pretransposed |
| 219 | bool B_is_pretransposed() const override { |
| 220 | return true; |
| 221 | } |
| 222 | |
| 223 | bool B_pretranspose_required() const override { |
| 224 | return (_B_transposed==nullptr); |
| 225 | } |
| 226 | |
| 227 | size_t get_B_pretransposed_array_size() const override { |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 228 | return roundup(_Nsize, strategy::out_width()) * roundup(_Ksize, strategy::k_unroll()) * _nmulti * sizeof(Toi); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 229 | } |
| 230 | |
| 231 | void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override { |
| 232 | Toi *buffer = reinterpret_cast<Toi *>(in_buffer); |
| 233 | _B_transposed = buffer; |
| 234 | strategy strat(_ci); |
| 235 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 236 | for (unsigned int multi=0; multi<_nmulti; multi++) { |
| 237 | for (unsigned int k0=0; k0<_Ksize; k0+=_k_block) { |
| 238 | const unsigned int kmax = std::min(k0 + _k_block, _Ksize); |
| 239 | const unsigned int k_size = roundup(kmax-k0, strategy::k_unroll()); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 240 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 241 | for (unsigned int x0=0; x0<_Nsize; x0+=_n_block) { |
| 242 | const unsigned int xmax = std::min(x0+_n_block, _Nsize); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 243 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 244 | const unsigned int size = roundup(xmax-x0, strategy::out_width()) * k_size; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 245 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 246 | strat.transforms.PrepareB( buffer, B + (multi * B_multi_stride), ldb, |
| 247 | x0, xmax, k0, kmax, _trB); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 248 | |
Georgios Pinitas | 1d48065 | 2019-01-23 11:24:50 +0000 | [diff] [blame] | 249 | buffer += size; |
| 250 | } |
| 251 | } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 252 | } |
| 253 | } |
| 254 | |
| 255 | void set_pretransposed_B_data(void *in_buffer) override { |
| 256 | _B_transposed = reinterpret_cast<Toi *>(in_buffer); |
| 257 | } |
| 258 | }; |
| 259 | |
| 260 | } // namespace arm_gemm |