Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 1 | /* |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [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 | */ |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 24 | #include "GEMMLowp.h" |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/Types.h" |
Georgios Pinitas | 5a7e776 | 2017-12-01 16:27:29 +0000 | [diff] [blame] | 27 | #include "tests/validation/reference/UtilsQuantizedAsymm.h" |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 28 | |
| 29 | #include <limits> |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace test |
| 34 | { |
| 35 | namespace validation |
| 36 | { |
| 37 | namespace reference |
| 38 | { |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 39 | namespace |
| 40 | { |
| 41 | template <typename T> |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 42 | struct DataTypeExtractor |
| 43 | { |
| 44 | static DataType data_type() |
| 45 | { |
| 46 | DataType data_type = DataType::UNKNOWN; |
| 47 | if(std::is_same<T, int8_t>::value) |
| 48 | { |
| 49 | data_type = DataType::QASYMM8_SIGNED; |
| 50 | } |
| 51 | else if(std::is_same<T, uint8_t>::value) |
| 52 | { |
| 53 | data_type = DataType::QASYMM8; |
| 54 | } |
| 55 | else if(std::is_same<T, int16_t>::value) |
| 56 | { |
| 57 | data_type = DataType::QSYMM16; |
| 58 | } |
| 59 | return data_type; |
| 60 | } |
| 61 | }; |
| 62 | |
| 63 | template <typename T> |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 64 | void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<uint8_t> *dst, int32_t result_offset, std::vector<int32_t> result_mult_int, |
| 65 | std::vector<int32_t> result_shift, int32_t min, int32_t max) |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 66 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 67 | const int cols_in = in->shape().x(); |
| 68 | const bool is_per_channel = result_mult_int.size() > 1; |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 69 | |
| 70 | for(int i = 0; i < in->num_elements(); ++i) |
| 71 | { |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 72 | int32_t result = ((*in)[i] + result_offset); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 73 | |
| 74 | if(bias != nullptr) |
| 75 | { |
| 76 | result += (*bias)[i % cols_in]; |
| 77 | } |
| 78 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 79 | result *= (is_per_channel) ? result_mult_int[i % cols_in] : result_mult_int[0]; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 80 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 81 | result >>= (is_per_channel) ? result_shift[i % cols_in] : result_shift[0]; |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 82 | |
| 83 | // Bounded ReLu |
| 84 | if(min != max) |
| 85 | { |
| 86 | result = std::max(min, std::min(max, result)); |
| 87 | } |
| 88 | |
| 89 | (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result))); |
| 90 | } |
| 91 | } |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 92 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 93 | template <typename TIn, typename TOut> |
| 94 | void quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> *in, const SimpleTensor<TIn> *bias, SimpleTensor<TOut> *dst, std::vector<int32_t> result_fixedpoint_multiplier, |
| 95 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max) |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 96 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 97 | const int cols_in = in->shape().x(); |
| 98 | const bool is_per_channel = result_fixedpoint_multiplier.size() > 1; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 99 | |
| 100 | for(int i = 0; i < in->num_elements(); ++i) |
| 101 | { |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 102 | TIn result = (*in)[i]; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 103 | |
| 104 | if(bias != nullptr) |
| 105 | { |
| 106 | result += (*bias)[i % cols_in]; |
| 107 | } |
| 108 | |
| 109 | // Fixed point multiplication |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 110 | const int32_t multiplier = (is_per_channel) ? result_fixedpoint_multiplier[i % cols_in] : result_fixedpoint_multiplier[0]; |
| 111 | const int32_t shift = (is_per_channel) ? result_shift[i % cols_in] : result_shift[0]; |
| 112 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 113 | if(shift < 0) |
| 114 | { |
| 115 | result = asymm_int_mult(result * (1 << (-shift)), multiplier); |
| 116 | } |
| 117 | else |
| 118 | { |
| 119 | result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, multiplier), shift); |
| 120 | } |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 121 | result += result_offset_after_shift; |
| 122 | |
| 123 | // Bounded ReLu |
| 124 | if(min != max) |
| 125 | { |
| 126 | result = std::max(min, std::min(max, result)); |
| 127 | } |
| 128 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 129 | (*dst)[i] = static_cast<TOut>(std::max<TIn>(std::numeric_limits<TOut>::lowest(), |
| 130 | std::min<TIn>(std::numeric_limits<TOut>::max(), result))); |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 131 | } |
| 132 | } |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 133 | } // namespace |
| 134 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 135 | template <typename T_out, typename T_in, typename T_in_1> |
| 136 | SimpleTensor<T_out> gemmlowp_matrix_multiply_core(const SimpleTensor<T_in> &a, const SimpleTensor<T_in_1> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 137 | { |
Michalis Spyrou | f3dfa27 | 2017-11-21 17:52:12 +0000 | [diff] [blame] | 138 | static_assert(std::is_same<typename std::decay<T_out>::type, int32_t>::value, "Only int32_t is allowed for the output"); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 139 | |
Michalis Spyrou | f3dfa27 | 2017-11-21 17:52:12 +0000 | [diff] [blame] | 140 | DataType dt = std::is_same<T_out, int32_t>::value ? DataType::S32 : DataType::U32; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 141 | SimpleTensor<T_out> c(shape_c, dt); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 142 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 143 | const int K = a.shape().x(); |
| 144 | const int M = a.shape().y(); |
| 145 | const int N = b.shape().x(); |
| 146 | const int D = a.shape().z(); // Number of matrices in a batch |
| 147 | |
| 148 | const int a_stride_z = K * M; |
| 149 | // Do not slide the matrix B along the 3rd dimension in case matrix B has less than 3 dimensions |
| 150 | const int b_stride_z = b.shape().num_dimensions() > 2 ? N * K : 0; |
| 151 | const int c_stride_z = N * M; |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 152 | |
Michalis Spyrou | f3dfa27 | 2017-11-21 17:52:12 +0000 | [diff] [blame] | 153 | std::vector<T_out> acc; |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 154 | acc.resize(N); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 155 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 156 | for(int depth = 0; depth < D; ++depth) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 157 | { |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 158 | const int base_addr_a = depth * a_stride_z; |
| 159 | const int base_addr_b = depth * b_stride_z; |
| 160 | const int base_addr_c = depth * c_stride_z; |
| 161 | |
| 162 | for(int i = 0; i < M; ++i) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 163 | { |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 164 | for(int j = 0; j < N; ++j) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 165 | { |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 166 | acc[j] = 0; |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 167 | } |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 168 | for(int k = 0; k < K; ++k) |
| 169 | { |
| 170 | const T_out tmp_a = a_offset + static_cast<T_out>(a[base_addr_a + k + i * K]); |
| 171 | for(int j = 0; j < N; ++j) |
| 172 | { |
| 173 | const T_out tmp_b = b_offset + static_cast<T_out>(b[base_addr_b + j + k * N]); |
| 174 | const T_out mult_as_int = tmp_a * tmp_b; |
| 175 | acc[j] += mult_as_int; |
| 176 | } |
| 177 | } |
| 178 | for(int j = 0; j < N; ++j) |
| 179 | { |
| 180 | c[base_addr_c + j + i * N] = acc[j]; |
| 181 | } |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 182 | } |
| 183 | } |
| 184 | |
| 185 | return c; |
| 186 | } |
| 187 | |
Pablo Tello | 181e651 | 2017-11-15 13:28:27 +0000 | [diff] [blame] | 188 | // used to validate assembly kernels which don't know anything about offsets |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 189 | template <typename T1, typename T2, typename T3> |
| 190 | SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T3> &b, TensorShape shape_c) |
Pablo Tello | 181e651 | 2017-11-15 13:28:27 +0000 | [diff] [blame] | 191 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 192 | return gemmlowp_matrix_multiply_core<T1, T2, T3>(a, b, shape_c, 0, 0); |
Pablo Tello | 181e651 | 2017-11-15 13:28:27 +0000 | [diff] [blame] | 193 | } |
| 194 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 195 | template <typename T> |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 196 | SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, std::vector<int32_t> result_mult_int, std::vector<int32_t> result_shift, |
| 197 | int32_t min, int32_t max) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 198 | { |
| 199 | SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8); |
| 200 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 201 | quantize_down_int32_to_uint8_scale<T>(&in, nullptr, &dst, result_offset, result_mult_int, result_shift, min, max); |
| 202 | |
| 203 | return dst; |
| 204 | } |
| 205 | |
| 206 | template <typename T> |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 207 | SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_offset, std::vector<int32_t> result_mult_int, |
| 208 | std::vector<int32_t> result_shift, int32_t min, int32_t max) |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 209 | { |
| 210 | SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8); |
| 211 | |
| 212 | quantize_down_int32_to_uint8_scale<T>(&in, &bias, &dst, result_offset, result_mult_int, result_shift, min, max); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 213 | |
| 214 | return dst; |
| 215 | } |
| 216 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 217 | template <typename TIn, typename TOut> |
| 218 | SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, std::vector<int32_t> result_fixedpoint_multiplier, std::vector<int32_t> result_shift, |
| 219 | int32_t result_offset_after_shift, int32_t min, int32_t max) |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 220 | { |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 221 | SimpleTensor<TOut> dst(in.shape(), DataTypeExtractor<TOut>::data_type()); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 222 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 223 | quantize_down_scale_by_fixedpoint<TIn, TOut>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 224 | |
| 225 | return dst; |
| 226 | } |
| 227 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 228 | template <typename TIn, typename TOut> |
| 229 | SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, const SimpleTensor<TIn> &bias, std::vector<int32_t> result_fixedpoint_multiplier, |
| 230 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max) |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 231 | { |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 232 | SimpleTensor<TOut> dst(in.shape(), DataTypeExtractor<TOut>::data_type()); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 233 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 234 | quantize_down_scale_by_fixedpoint<TIn, TOut>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 235 | |
| 236 | return dst; |
| 237 | } |
| 238 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame^] | 239 | template SimpleTensor<uint8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier, |
| 240 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
| 241 | template SimpleTensor<uint8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, |
| 242 | std::vector<int32_t> result_fixedpoint_multiplier, |
| 243 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
| 244 | template SimpleTensor<int8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier, |
| 245 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
| 246 | template SimpleTensor<int8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, |
| 247 | std::vector<int32_t> result_fixedpoint_multiplier, |
| 248 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
| 249 | template SimpleTensor<int16_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier, |
| 250 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
| 251 | template SimpleTensor<int16_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, |
| 252 | std::vector<int32_t> result_fixedpoint_multiplier, |
| 253 | std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 254 | template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, std::vector<int32_t> result_mult_int, |
| 255 | std::vector<int32_t> result_shift, int32_t min, int32_t max); |
| 256 | template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_offset, std::vector<int32_t> result_mult_int, |
| 257 | std::vector<int32_t> result_shift, int32_t min, int32_t max); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 258 | template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset); |
| 259 | template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 260 | template SimpleTensor<int32_t> gemmlowp<int32_t, int8_t, int8_t>(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c); |
| 261 | template SimpleTensor<int32_t> gemmlowp<int32_t, uint8_t, uint8_t>(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, TensorShape shape_c); |
| 262 | template SimpleTensor<int32_t> gemmlowp<int32_t, uint8_t, int8_t>(const SimpleTensor<uint8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c); |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 263 | } // namespace reference |
| 264 | } // namespace validation |
| 265 | } // namespace test |
| 266 | } // namespace arm_compute |