xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/quantize_wrapper.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2021 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 
25 #pragma once
26 
27 #include "arm_gemm.hpp"
28 
29 #include "barrier.hpp"
30 #include "gemm_implementation.hpp"
31 #include "quantized.hpp"
32 
33 namespace arm_gemm {
34 
35 /* Quantized wrapper - do an integer GEMM and wrap around the quantization. */
36 
37 template<typename To, typename Tr, typename Tgemm>
38 class QuantizeWrapper : public GemmCommon<To, Tr> {
39 private:
40     UniqueGemmCommon<To, Tgemm>  _subgemm = nullptr;
41     int32_t                     *_row_sums = nullptr;
42     int32_t                     *_col_sums = nullptr;
43     Requantize32                 _params;
44     GemmArgs                     _args;
45     barrier                      _barrier;
46 
47     void *working_space = nullptr;
48     bool  arrays_set = false;
49 
50     /* We need a subgemm which outputs the 32-bit intermediates - how much space is needed for that? */
subgemm_output_size() const51     size_t subgemm_output_size() const {
52         return (_args._Msize * _args._Nsize * _args._nbatches * _args._nmulti * sizeof(int32_t));
53     }
54 
col_sum_size() const55     size_t col_sum_size() const {
56         return (_args._Nsize * _args._nmulti * sizeof(int32_t));
57     }
58 
row_sum_size() const59     size_t row_sum_size() const {
60         return (_args._Msize * _args._nbatches * _args._nmulti * sizeof(int32_t));
61     }
62 
63     /* Local working space: We need space for the subgemm output (above) and
64      * the row sums.  */
local_working_size() const65     size_t local_working_size() const {
66         return subgemm_output_size() + row_sum_size();
67     }
68 
set_child_arrays()69     void set_child_arrays() {
70         if (working_space == nullptr || arrays_set == false)
71             return;
72 
73         /* Use the first part of our working space for the subgemm result, pass the operand details straight through. */
74         _subgemm->set_arrays(this->_Aptr, this->_lda, this->_A_batch_stride, this->_A_multi_stride,
75                              this->_Bptr, this->_ldb,                        this->_B_multi_stride,
76                              reinterpret_cast<Tgemm *>(working_space), _args._Nsize, (_args._Nsize * _args._Msize), (_args._Nsize * _args._Msize * _args._nbatches),
77                              nullptr, 0);
78     }
79 
col_sums_pretransposed(const To * B,const int ldb,const int B_multi_stride)80     void col_sums_pretransposed(const To *B, const int ldb, const int B_multi_stride) {
81         for (unsigned int multi=0; multi<_args._nmulti; multi++) {
82             compute_col_sums(_params, _args._Nsize, _args._Ksize, B + (multi * B_multi_stride), ldb, _col_sums + (multi * _args._Nsize), _args._Ksize, multi, 0);
83         }
84     }
85 
requantize_runtime(unsigned int threadid)86     void requantize_runtime(unsigned int threadid) {
87         unsigned int first_row = (threadid * _args._Msize) / _args._maxthreads;
88         unsigned int last_row = ((threadid+1) * _args._Msize) / _args._maxthreads;
89 
90         for (unsigned int multi=0; multi<_args._nmulti; multi++) {
91             for (unsigned int batch=0; batch<_args._nbatches; batch++) {
92                 /* Compute row sums now */
93                 compute_row_sums(_params, _args._Ksize, (last_row - first_row), this->_Aptr + (multi * this->_A_multi_stride) + (batch * this->_A_batch_stride) + (first_row * this->_lda),
94                                            this->_lda, _row_sums + (multi * _args._nbatches * _args._Msize) + (batch * _args._Msize) + first_row);
95                 // If we don't care about negative values, call the version of this function that doesn't correct before shifting.
96                 // 'c_offset' represents zero, so if the lowest possible quantized output value is the same or more than that we will not output negative numbers.
97                 requantize_block_32(_params, _args._Nsize, (last_row - first_row),
98                                     reinterpret_cast<Tgemm *>(working_space) + (multi * (_args._Msize * _args._Nsize * _args._nbatches)) + (batch * (_args._Msize * _args._Nsize)) + (first_row * _args._Nsize),
99                                     _args._Nsize,
100                                     this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (first_row * this->_ldc), this->_ldc,
101                                     _row_sums + (multi * _args._nbatches * _args._Msize) + (batch * _args._Msize) + first_row,
102                                     _col_sums + (multi * _args._Nsize), 0);
103             }
104         }
105     }
106 
107 
108 public:
109     QuantizeWrapper(const QuantizeWrapper &) = delete;
110     QuantizeWrapper operator=(const QuantizeWrapper &) = delete;
111 
QuantizeWrapper(const GemmArgs & args,const Requantize32 & qp)112     QuantizeWrapper(const GemmArgs &args, const Requantize32 &qp) : _params(qp), _args(args), _barrier(args._maxthreads) {
113         GemmArgs newargs = GemmArgs(args._ci, args._Msize, args._Nsize, args._Ksize, args._Ksections, args._nbatches, args._nmulti, args._indirect_input, Activation(), args._maxthreads);
114         _subgemm = gemm<To, Tgemm>(newargs);
115 
116         if (_subgemm == nullptr) {
117             return;
118         }
119     }
120 
set_arrays(const To * A,const int lda,const int A_batch_stride,const int A_multi_stride,const To * B,const int ldb,const int B_multi_stride,Tr * C,const int ldc,const int C_batch_stride,const int C_multi_stride,const Tr * bias,const int bias_multi_stride)121     void set_arrays(const To *A, const int lda, const int A_batch_stride, const int A_multi_stride,
122                     const To *B, const int ldb, const int B_multi_stride,
123                           Tr *C, const int ldc, const int C_batch_stride, const int C_multi_stride,
124                     const Tr *bias, const int bias_multi_stride) override {
125         GemmCommon<To, Tr>::set_arrays(A, lda, A_batch_stride, A_multi_stride, B, ldb, B_multi_stride, C, ldc, C_batch_stride, C_multi_stride, bias, bias_multi_stride);
126 
127         arrays_set = true;
128         set_child_arrays();
129     }
130 
get_window_size() const131     ndrange_t get_window_size() const override {
132         return { _subgemm->get_window_size() };
133     }
134 
set_nthreads(int nthreads)135     void set_nthreads(int nthreads) override {
136         _subgemm->set_nthreads(nthreads);
137         _barrier.set_nthreads(nthreads);
138         _args._maxthreads = nthreads;
139     }
140 
execute(const ndcoord_t & work_range,const ndcoord_t & thread_locator,int threadid)141     void execute(const ndcoord_t &work_range, const ndcoord_t &thread_locator, int threadid) override {
142         _subgemm->execute(work_range, thread_locator, threadid);
143 
144         _barrier.arrive_and_wait();
145 
146         requantize_runtime(threadid);
147     }
148 
get_working_size() const149     size_t get_working_size() const override {
150         return _subgemm->get_working_size() + local_working_size();
151     }
152 
153     // Space arrangement:
154 
155     // ptr
156     // V
157     // | subgemm output | row_sums | subgemm working space |
set_working_space(void * space)158     void set_working_space(void *space) override {
159         uintptr_t space_int = reinterpret_cast<uintptr_t>(space);
160 
161         working_space = space;
162         _subgemm->set_working_space(reinterpret_cast<void *>(space_int + local_working_size()));
163 
164         _row_sums = reinterpret_cast<int32_t *>(space_int + subgemm_output_size());
165 
166         set_child_arrays();
167     }
168 
B_is_pretransposed() const169     bool B_is_pretransposed() const override {
170         /* We clear this flag if the subgemm isn't pretransposed, so just return its value */
171         return _subgemm->B_is_pretransposed();
172     }
173 
B_pretranspose_required() const174     bool B_pretranspose_required() const override {
175         return _subgemm->B_pretranspose_required();
176     }
177 
get_B_pretransposed_array_size() const178     size_t get_B_pretransposed_array_size() const override {
179         return _subgemm->get_B_pretransposed_array_size() + col_sum_size();
180     }
181 
requantize_bias(void * in_buffer,const To * B,const int ldb,const int B_multi_stride)182     void requantize_bias(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
183         _col_sums = reinterpret_cast<int32_t *>(in_buffer);
184         col_sums_pretransposed(B, ldb, B_multi_stride);
185     }
186 
pretranspose_B_array(void * buffer,const To * B,const int ldb,const int B_multi_stride)187     void pretranspose_B_array(void *buffer, const To *B, const int ldb, const int B_multi_stride) override {
188         uintptr_t buffer_int = reinterpret_cast<uintptr_t>(buffer);
189         _subgemm->pretranspose_B_array(reinterpret_cast<void *>(buffer_int + col_sum_size()), B, ldb, B_multi_stride);
190 
191         requantize_bias(buffer, B, ldb, B_multi_stride);
192     }
193 
set_pretransposed_B_data(void * buffer)194     void set_pretransposed_B_data(void *buffer) override {
195         uintptr_t buffer_int = reinterpret_cast<uintptr_t>(buffer);
196         _subgemm->set_pretransposed_B_data(reinterpret_cast<void *>(buffer_int + col_sum_size()));
197         _col_sums = reinterpret_cast<int32_t *>(buffer);
198     }
199 
set_quantized_bias(const int32_t * bias,size_t bias_multi_stride)200     void set_quantized_bias(const int32_t *bias, size_t bias_multi_stride) override {
201         _params.bias = bias;
202         _params.bias_multi_stride = bias_multi_stride;
203     }
204 
get_config()205     GemmConfig get_config() override {
206         GemmConfig c = _subgemm->get_config();
207 
208         std::string n = "quantize_wrapper[";
209         n.append(c.filter);
210         n.append("]");
211 
212         c.method = GemmMethod::QUANTIZE_WRAPPER;
213         c.filter = n;
214 
215         return c;
216     }
217 };
218 
219 } // namespace arm_gemm
220