Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 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; |
Michalis Spyrou | 71ac903 | 2019-11-14 14:31:44 +0000 | [diff] [blame] | 43 | Requantize32 _params; |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 44 | GemmArgs _args; |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 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? */ |
| 51 | size_t subgemm_output_size() const { |
| 52 | return (_args._Msize * _args._Nsize * _args._nbatches * _args._nmulti * sizeof(int32_t)); |
| 53 | } |
| 54 | |
| 55 | size_t col_sum_size() const { |
| 56 | return (_args._Nsize * _args._nmulti * sizeof(int32_t)); |
| 57 | } |
| 58 | |
| 59 | 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. If the GEMM is not pretransposed we need to store the |
| 65 | * column sums here too. */ |
| 66 | size_t local_working_size() const { |
| 67 | size_t sz = subgemm_output_size() + row_sum_size(); |
| 68 | |
| 69 | if (_args._pretransposed_hint) { |
| 70 | return sz; |
| 71 | } |
| 72 | |
| 73 | return sz + col_sum_size(); |
| 74 | } |
| 75 | |
| 76 | void set_child_arrays() { |
| 77 | if (working_space == nullptr || arrays_set == false) |
| 78 | return; |
| 79 | |
| 80 | /* Use the first part of our working space for the subgemm result, pass the operand details straight through. */ |
| 81 | _subgemm->set_arrays(this->_Aptr, this->_lda, this->_A_batch_stride, this->_A_multi_stride, |
| 82 | this->_Bptr, this->_ldb, this->_B_multi_stride, |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 83 | reinterpret_cast<Tgemm *>(working_space), _args._Nsize, (_args._Nsize * _args._Msize), (_args._Nsize * _args._Msize * _args._nbatches), |
| 84 | nullptr, 0); |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 85 | } |
| 86 | |
| 87 | void col_sums_pretransposed(const To *B, const int ldb, const int B_multi_stride) { |
| 88 | for (unsigned int multi=0; multi<_args._nmulti; multi++) { |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 89 | compute_col_sums(_params, _args._Nsize, _args._Ksize, B + (multi * B_multi_stride), ldb, _col_sums + (multi * _args._Nsize), _args._Ksize, multi, 0); |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 90 | } |
| 91 | } |
| 92 | |
| 93 | void col_sums_runtime(unsigned int threadid) { |
| 94 | unsigned int first_col = (threadid * _args._Nsize) / _args._maxthreads; |
| 95 | unsigned int last_col = ((threadid + 1) * _args._Nsize) / _args._maxthreads; |
| 96 | |
| 97 | for (unsigned int multi=0; multi<_args._nmulti; multi++) { |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 98 | compute_col_sums(_params, (last_col - first_col), _args._Ksize, this->_Bptr + (multi * this->_B_multi_stride) + first_col, this->_ldb, _col_sums + (multi * _args._Nsize) + first_col, _args._Ksize, multi, first_col); |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 99 | } |
| 100 | } |
| 101 | |
| 102 | void requantize_runtime(unsigned int threadid) { |
| 103 | unsigned int first_row = (threadid * _args._Msize) / _args._maxthreads; |
| 104 | unsigned int last_row = ((threadid+1) * _args._Msize) / _args._maxthreads; |
| 105 | |
| 106 | for (unsigned int multi=0; multi<_args._nmulti; multi++) { |
| 107 | for (unsigned int batch=0; batch<_args._nbatches; batch++) { |
| 108 | /* Compute row sums now */ |
| 109 | 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), |
| 110 | this->_lda, _row_sums + (multi * _args._nbatches * _args._Msize) + (batch * _args._Msize) + first_row); |
| 111 | // If we don't care about negative values, call the version of this function that doesn't correct before shifting. |
| 112 | // '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. |
| 113 | requantize_block_32(_params, _args._Nsize, (last_row - first_row), |
| 114 | reinterpret_cast<Tgemm *>(working_space) + (multi * (_args._Msize * _args._Nsize * _args._nbatches)) + (batch * (_args._Msize * _args._Nsize)) + (first_row * _args._Nsize), |
| 115 | _args._Nsize, |
| 116 | this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (first_row * this->_ldc), this->_ldc, |
| 117 | _row_sums + (multi * _args._nbatches * _args._Msize) + (batch * _args._Msize) + first_row, |
| 118 | _col_sums + (multi * _args._Nsize)); |
| 119 | } |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | |
| 124 | public: |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 125 | QuantizeWrapper(const QuantizeWrapper &) = delete; |
| 126 | QuantizeWrapper operator=(const QuantizeWrapper &) = delete; |
| 127 | |
Michalis Spyrou | 71ac903 | 2019-11-14 14:31:44 +0000 | [diff] [blame] | 128 | QuantizeWrapper(const GemmArgs &args, const Requantize32 &qp) : _params(qp), _args(args), _barrier(args._maxthreads) { |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 129 | GemmArgs newargs = GemmArgs(args._ci, args._Msize, args._Nsize, args._Ksize, args._nbatches, args._nmulti, args._trA, args._trB, Activation(), args._maxthreads, args._pretransposed_hint, nullptr); |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 130 | _subgemm = gemm<To, Tgemm>(newargs); |
| 131 | |
| 132 | if (_subgemm == nullptr) { |
| 133 | return; |
| 134 | } |
| 135 | |
| 136 | if (!_subgemm->B_is_pretransposed()) { |
| 137 | _args._pretransposed_hint = false; |
| 138 | } |
| 139 | } |
| 140 | |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 141 | void set_arrays(const To *A, const int lda, const int A_batch_stride, const int A_multi_stride, |
| 142 | const To *B, const int ldb, const int B_multi_stride, |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 143 | Tr *C, const int ldc, const int C_batch_stride, const int C_multi_stride, |
| 144 | const Tr *bias, const int bias_multi_stride) override { |
| 145 | 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); |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 146 | |
| 147 | arrays_set = true; |
| 148 | set_child_arrays(); |
| 149 | } |
| 150 | |
| 151 | unsigned int get_window_size() const override { |
| 152 | return _subgemm->get_window_size(); |
| 153 | } |
| 154 | |
| 155 | void set_nthreads(int nthreads) override { |
| 156 | _subgemm->set_nthreads(nthreads); |
| 157 | _barrier.set_nthreads(nthreads); |
| 158 | _args._maxthreads = nthreads; |
| 159 | } |
| 160 | |
| 161 | void execute(unsigned int start, unsigned int end, int threadid) override { |
| 162 | _subgemm->execute(start, end, threadid); |
| 163 | if (!_args._pretransposed_hint) { |
| 164 | col_sums_runtime(threadid); |
| 165 | } |
| 166 | |
| 167 | _barrier.arrive_and_wait(); |
| 168 | |
| 169 | requantize_runtime(threadid); |
| 170 | } |
| 171 | |
| 172 | size_t get_working_size() const override { |
| 173 | return _subgemm->get_working_size() + local_working_size(); |
| 174 | } |
| 175 | |
| 176 | // Space arrangement: |
| 177 | |
| 178 | // ptr |
| 179 | // V |
| 180 | // | subgemm output | row_sums | col_sums (if not pretransposed | subgemm working space | |
| 181 | void set_working_space(void *space) override { |
| 182 | uintptr_t space_int = reinterpret_cast<uintptr_t>(space); |
| 183 | |
| 184 | working_space = space; |
| 185 | _subgemm->set_working_space(reinterpret_cast<void *>(space_int + local_working_size())); |
| 186 | |
| 187 | _row_sums = reinterpret_cast<int32_t *>(space_int + subgemm_output_size()); |
| 188 | if (!_args._pretransposed_hint) { |
| 189 | _col_sums = reinterpret_cast<int32_t *>(space_int + subgemm_output_size() + row_sum_size()); |
| 190 | } |
| 191 | |
| 192 | set_child_arrays(); |
| 193 | } |
| 194 | |
| 195 | bool B_is_pretransposed() const override { |
| 196 | /* We clear this flag if the subgemm isn't pretransposed, so just return its value */ |
| 197 | return _args._pretransposed_hint; |
| 198 | } |
| 199 | |
| 200 | bool B_pretranspose_required() const override { |
| 201 | return _subgemm->B_pretranspose_required(); |
| 202 | } |
| 203 | |
| 204 | size_t get_B_pretransposed_array_size() const override { |
| 205 | if (_args._pretransposed_hint) { |
| 206 | return _subgemm->get_B_pretransposed_array_size() + col_sum_size(); |
| 207 | } |
| 208 | |
| 209 | return 0; |
| 210 | } |
| 211 | |
| 212 | void pretranspose_B_array(void *buffer, const To *B, const int ldb, const int B_multi_stride) override { |
| 213 | if (!_args._pretransposed_hint) { |
| 214 | return; |
| 215 | } |
| 216 | |
| 217 | uintptr_t buffer_int = reinterpret_cast<uintptr_t>(buffer); |
| 218 | _subgemm->pretranspose_B_array(reinterpret_cast<void *>(buffer_int + col_sum_size()), B, ldb, B_multi_stride); |
| 219 | |
| 220 | _col_sums = reinterpret_cast<int32_t *>(buffer); |
| 221 | |
| 222 | col_sums_pretransposed(B, ldb, B_multi_stride); |
| 223 | } |
| 224 | |
| 225 | void set_pretransposed_B_data(void *buffer) override { |
| 226 | if (!_args._pretransposed_hint) { |
| 227 | return; |
| 228 | } |
| 229 | |
| 230 | uintptr_t buffer_int = reinterpret_cast<uintptr_t>(buffer); |
| 231 | _subgemm->set_pretransposed_B_data(reinterpret_cast<void *>(buffer_int + col_sum_size())); |
| 232 | _col_sums = reinterpret_cast<int32_t *>(buffer); |
| 233 | } |
| 234 | |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 235 | void set_quantized_bias(const int32_t *bias, size_t bias_multi_stride) override { |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 236 | _params.bias = bias; |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 237 | _params.bias_multi_stride = bias_multi_stride; |
Georgios Pinitas | cfa2bba | 2019-06-27 17:00:52 +0100 | [diff] [blame] | 238 | } |
| 239 | }; |
| 240 | |
| 241 | } // namespace arm_gemm |