xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/gemm_hybrid.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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 <algorithm>
27 #include <cassert>
28 
29 #include "arm_gemm.hpp"
30 #include "bias_adder.hpp"
31 #include "ndrange.hpp"
32 #include "performance_parameters.hpp"
33 #include "transform.hpp"
34 #include "utils.hpp"
35 
36 #ifdef CYCLE_PROFILING
37 #include "profiler.hpp"
38 #endif
39 
40 namespace arm_gemm {
41 
42 // Implementation of the GemmCommon abstract class.
43 template<typename strategy, typename To, typename Tr>
44 class GemmHybrid : public GemmCommon<To, Tr> {
45     typedef typename strategy::operand_type Toi;
46     typedef typename strategy::result_type Tri;
47 
48     /* const properties set by constructor */
49     const CPUInfo * const _ci;
50 
51     const unsigned int _Msize;
52     const unsigned int _Nsize;
53     const unsigned int _Ksize;
54 
55     const unsigned int _nbatches;
56     const unsigned int _nmulti;
57 
58     const Activation _act;
59 
60     /* Blocking info */
61     const unsigned int _k_block;
62     const unsigned int _n_block;
63     const unsigned int _Mround;
64 
65     /* Pretransposed buffer. */
66     const Toi *_B_transposed=nullptr;
67 
68     const NDRange<4> _window_range;
69 
compute_k_block(const GemmArgs & args)70     static unsigned int compute_k_block(const GemmArgs &args) {
71         // Some kernels don't support accumulate mode - these can't do K blocking at all.
72         if (!strategy::supports_accumulate()) {
73             return args._Ksize;
74         }
75 
76         if (args._cfg && args._cfg->inner_block_size) {
77             return roundup(args._cfg->inner_block_size, strategy::k_unroll());
78         }
79 
80         // Target block size (512 for FP32, scaling for other types).  Don't block until size reaches 1.5X this.
81         unsigned int target_block_size = 2048 / sizeof(To);
82 
83         if (args._Ksize >= ((3 * target_block_size) / 2)) {
84             unsigned int target_blocks = iceildiv(args._Ksize, target_block_size);
85 
86             unsigned int block_size = iceildiv(args._Ksize, target_blocks);
87 
88             block_size = roundup(block_size, strategy::k_unroll());
89 
90             return block_size;
91         }
92 
93         return args._Ksize;
94     }
95 
96     // New N blocking strategy: if it's narrow, or much taller than it is wide, do the full width.  Otherwise do a
97     // single block.
compute_n_block(const GemmArgs & args)98     static unsigned int compute_n_block(const GemmArgs &args) {
99         if (args._cfg && args._cfg->outer_block_size) {
100             unsigned int n_block = args._cfg->outer_block_size;
101 
102             // Needs to be (at least a single) multiple of the kernel output width.
103             n_block /= strategy::out_width();
104             n_block = std::max(n_block, 1u) * strategy::out_width();
105 
106             return n_block;
107         }
108 
109         if (args._Nsize <= 64) {
110             return args._Nsize;
111         }
112 
113         if ((args._Msize / args._Nsize) > 155) {
114             return args._Nsize;
115         }
116 
117         // Go slightly wider if thread count and depth are small.
118         if ((args._Ksize <= 128) && (args._maxthreads <= 16)) {
119             return strategy::out_width() * 3;
120         }
121 
122         return strategy::out_width();
123     }
124 
125 public:
126     GemmHybrid(GemmHybrid &) = delete;
127     GemmHybrid & operator= (GemmHybrid &) = delete;
128 
129     /* Constructor */
GemmHybrid(const GemmArgs & args)130     GemmHybrid(const GemmArgs &args)
131               : _ci(args._ci), _Msize(args._Msize), _Nsize(args._Nsize), _Ksize(args._Ksize),
132                 _nbatches(args._nbatches), _nmulti(args._nmulti),
133                 _act(args._act),
134                 _k_block(compute_k_block(args)), _n_block(compute_n_block(args)),
135                 _Mround(roundup(args._Msize, strategy::out_height())),
136                 _window_range(iceildiv(args._Msize, strategy::out_height()), _nbatches, iceildiv(_Nsize, _n_block), _nmulti) { }
137 
138     // Interface implementation - Compulsory functions
get_window_size() const139     ndrange_t get_window_size() const override {
140         return { _window_range.total_size() };
141     }
142 
143     // This kernel can always be dynamically scheduled.
supports_dynamic_scheduling() const144     bool supports_dynamic_scheduling() const override {
145         return true;
146     }
147 
148     // Execute
execute(const ndcoord_t & work_range,const ndcoord_t &,int)149     void execute(const ndcoord_t &work_range, const ndcoord_t &, int) override {
150 #ifdef CYCLE_PROFILING
151         profiler prof;
152 #endif
153         strategy strat(_ci);
154 
155         /* Make sure we've been set up correctly. */
156         assert(_B_transposed);
157         static_assert(std::is_same<To, Toi>::value, "gemm_native: Operand types must be the same.");
158         static_assert(std::is_same<Tr, Tri>::value, "gemm_native: Result types must be the same.");
159 
160         /* For now, each work item implies all the K for a given output
161          * pixel (so we don't need to synchronize access to the output
162          * array).  So separate the loop over K blocks here.  */
163         for (unsigned int k0=0; k0<_Ksize; k0+=_k_block) {
164             unsigned int kmax   = std::min(k0 + _k_block, _Ksize);
165             unsigned int kern_k = roundup(kmax-k0, strategy::k_unroll());
166 
167             const bool first_pass = (k0 == 0);
168             const bool last_pass = (kmax == _Ksize);
169 
170             auto p = _window_range.iterator(work_range.get_position(0), work_range.get_position_end(0));
171 
172             if (p.done()) {
173                 return;
174             }
175 
176             do {
177                 const unsigned int m_start = p.dim(0) * strategy::out_height();
178                 const unsigned int m_end   = std::min(p.dim0_max() * strategy::out_height(), _Msize);
179                 const unsigned int batch   = p.dim(1);
180                 const unsigned int n0      = p.dim(2) * _n_block;
181                 const unsigned int nmax    = std::min(n0 + _n_block, _Nsize);
182                 const unsigned int multi   = p.dim(3);
183 
184                 const Toi *b_panel = _B_transposed +
185                                      (multi * roundup(_Nsize, strategy::out_width()) * roundup(_Ksize, strategy::k_unroll())) +
186                                      (k0 * roundup(_Nsize, strategy::out_width())) +
187                                      (n0 * kern_k);
188 
189 #ifdef CYCLE_PROFILING
190                 auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(m_end - m_start) * kern_k * roundup(nmax-n0, strategy::out_width()));
191 #endif
192 
193                 strat.kernel(this->_Aptr + (multi * this->_A_multi_stride) + (batch * this->_A_batch_stride) + (m_start * this->_lda) + k0, this->_lda,
194                              b_panel,
195                              this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (m_start * this->_ldc) + n0, this->_ldc,
196                              (m_end - m_start), (nmax - n0), kmax-k0,
197                              (strategy::supports_bias() && first_pass && this->_bias) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr,
198                              last_pass ? _act : Activation(), !first_pass);
199 
200                 // Add bias externally if needed
201                 if (!strategy::supports_bias() && this->_bias && first_pass) {
202                     bias_adder(this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (m_start * this->_ldc) + n0, this->_ldc,
203                                this->_bias + (multi * this->_bias_multi_stride) + n0,
204                                (m_end - m_start), (nmax - n0));
205                 }
206 
207             } while (p.next_dim1());
208         }
209     }
210 
211     // Interface implementation - pretransposed
B_is_pretransposed() const212     bool B_is_pretransposed() const override {
213         return true;
214     }
215 
B_pretranspose_required() const216     bool B_pretranspose_required() const override {
217         return (_B_transposed==nullptr);
218     }
219 
get_B_pretransposed_array_size() const220     size_t get_B_pretransposed_array_size() const override {
221         return roundup(_Nsize, strategy::out_width()) * roundup(_Ksize, strategy::k_unroll()) * _nmulti * sizeof(Toi);
222     }
223 
pretranspose_B_array(void * in_buffer,const To * B,const int ldb,const int B_multi_stride)224     void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
225         Toi *buffer = reinterpret_cast<Toi *>(in_buffer);
226         _B_transposed = buffer;
227         strategy strat(_ci);
228 
229         for (unsigned int multi=0; multi<_nmulti; multi++) {
230             for (unsigned int k0=0; k0<_Ksize; k0+=_k_block) {
231                 const unsigned int kmax = std::min(k0 + _k_block, _Ksize);
232                 const unsigned int k_size = roundup(kmax-k0, strategy::k_unroll());
233 
234                 for (unsigned int x0=0; x0<_Nsize; x0+=_n_block) {
235                     const unsigned int xmax = std::min(x0+_n_block, _Nsize);
236 
237                     const unsigned int size = roundup(xmax-x0, strategy::out_width()) * k_size;
238 
239                     strat.transforms.PrepareB( buffer, B + (multi * B_multi_stride), ldb,
240                                                x0, xmax, k0, kmax);
241 
242                     buffer += size;
243                 }
244             }
245         }
246     }
247 
set_pretransposed_B_data(void * in_buffer)248     void set_pretransposed_B_data(void *in_buffer) override {
249         _B_transposed = reinterpret_cast<Toi *>(in_buffer);
250     }
251 
252     // Estimate cycles for given problem given provided parameters
estimate_cycles(const GemmArgs & args,const PerformanceParameters & params)253     static uint64_t estimate_cycles(const GemmArgs &args, const PerformanceParameters &params) {
254         // Note: Current hybrid kernels don't actually round up height (they
255         // have paths for each possible height).  Might need to make this
256         // configurable in future.
257         uint64_t total_macs = static_cast<uint64_t>(args._nbatches) * args._nmulti * args._Msize * roundup(args._Nsize, strategy::out_width()) * roundup(args._Ksize, strategy::k_unroll());
258 
259         float mac_cycles = static_cast<float>(total_macs) / params.kernel_macs_cycle;
260 
261         // TODO: A bit of a kludge here: current hybrid kernels incur extra
262         // overhead where the width is not a multiple of kernel width.  It's
263         // most noticable where the overall width is quite low, so add 15%
264         // penalty for such widths.
265         if ((args._Nsize < strategy::out_width()) || (args._Nsize > strategy::out_width() && args._Nsize < 2*strategy::out_width())) {
266             mac_cycles *= 1.15f;
267         }
268 
269         uint64_t total_cycles = mac_cycles;
270 
271         return total_cycles;
272     }
273 
get_config()274     GemmConfig get_config() override {
275         GemmConfig c;
276 
277         c.method = GemmMethod::GEMM_HYBRID;
278         c.inner_block_size = _k_block;
279         c.outer_block_size = _n_block;
280         c.filter = get_type_name<strategy>();
281 
282         return c;
283     }
284 };
285 
286 } // namespace arm_gemm
287