Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 1 | /* |
Gunes Bayir | 9b921be | 2022-07-28 17:44:00 +0100 | [diff] [blame] | 2 | * Copyright (c) 2020-2022 Arm Limited. |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +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 | */ |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 24 | |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 25 | #include "src/cpu/kernels/add/generic/neon/impl.h" |
| 26 | #include "arm_compute/core/Helpers.h" |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 27 | #include "arm_compute/core/utils/misc/Traits.h" |
| 28 | #include "src/core/NEON/wrapper/wrapper.h" |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 29 | namespace arm_compute |
| 30 | { |
| 31 | namespace cpu |
| 32 | { |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 33 | template <typename ScalarType> |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 34 | void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 35 | { |
Michele Di Giorgio | 33f41fa | 2021-03-09 14:09:08 +0000 | [diff] [blame] | 36 | /** SIMD vector tag type. */ |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 37 | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<ScalarType, wrapper::traits::BitWidth::W128>; |
| 38 | |
| 39 | // Create input windows |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 40 | Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); |
| 41 | Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 42 | |
| 43 | // Clear X Dimension on execution window as we handle manually |
| 44 | Window win = window; |
| 45 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 46 | |
| 47 | constexpr int window_step_x = 16 / sizeof(ScalarType); |
| 48 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 49 | const auto window_end_x = static_cast<int>(window.x().end()); |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 50 | const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x(); |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 51 | |
| 52 | if(is_broadcast_across_x) |
| 53 | { |
| 54 | const bool is_broadcast_input_2 = input2_win.x().step() == 0; |
| 55 | Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; |
| 56 | Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 57 | const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0; |
| 58 | const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0; |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 59 | |
| 60 | // Clear X Dimension on execution window as we handle manually |
| 61 | non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 62 | |
| 63 | Iterator broadcast_input(broadcast_tensor, broadcast_win); |
| 64 | Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 65 | Iterator output(dst, win); |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 66 | |
| 67 | execute_window_loop(win, [&](const Coordinates &) |
| 68 | { |
| 69 | const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr()); |
| 70 | const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); |
| 71 | |
| 72 | const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr()); |
| 73 | const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{}); |
| 74 | |
| 75 | // Compute S elements per iteration |
| 76 | int x = window_start_x; |
| 77 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 78 | { |
| 79 | const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x); |
| 80 | const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v); |
| 81 | wrapper::vstore(output_ptr + x, res); |
| 82 | } |
| 83 | |
| 84 | // Compute left-over elements |
| 85 | for(; x < window_end_x; ++x) |
| 86 | { |
| 87 | const auto non_broadcast_v = *(non_broadcast_input_ptr + x); |
| 88 | *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v; |
| 89 | } |
| 90 | }, |
| 91 | broadcast_input, non_broadcast_input, output); |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | // Clear X Dimension on execution window as we handle manually |
| 96 | input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 97 | input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 98 | |
Sheri Zhang | 6124390 | 2021-01-12 18:25:16 +0000 | [diff] [blame] | 99 | Iterator input1(src0, input1_win); |
| 100 | Iterator input2(src1, input2_win); |
| 101 | Iterator output(dst, win); |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 102 | |
| 103 | execute_window_loop(win, [&](const Coordinates &) |
| 104 | { |
| 105 | const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); |
| 106 | const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); |
| 107 | const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); |
| 108 | |
| 109 | // Compute S elements per iteration |
| 110 | int x = window_start_x; |
| 111 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 112 | { |
| 113 | const auto val1 = wrapper::vloadq(input1_ptr + x); |
| 114 | const auto val2 = wrapper::vloadq(input2_ptr + x); |
| 115 | const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2); |
| 116 | wrapper::vstore(output_ptr + x, res); |
| 117 | } |
| 118 | |
| 119 | // Compute left-over elements |
| 120 | for(; x < window_end_x; ++x) |
| 121 | { |
| 122 | const auto val1 = *(input1_ptr + x); |
| 123 | const auto val2 = *(input2_ptr + x); |
| 124 | *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2; |
| 125 | } |
| 126 | }, |
| 127 | input1, input2, output); |
| 128 | } |
| 129 | } |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 130 | |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 131 | bool add_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst) |
| 132 | { |
| 133 | const auto iq0 = src0->quantization_info().uniform(); |
| 134 | const auto iq1 = src1->quantization_info().uniform(); |
| 135 | const auto oq = dst->quantization_info().uniform(); |
| 136 | |
| 137 | const auto scale0 = iq0.scale / oq.scale; |
| 138 | const auto scale1 = iq1.scale / oq.scale; |
| 139 | |
| 140 | if(scale0 < -31.f || scale0 > 31.f || scale1 < -31.f || scale1 > 31.f) |
| 141 | { |
| 142 | // The scale factor cannot be stored as 6.10 signed fixed-point number. |
| 143 | return false; |
| 144 | } |
| 145 | |
| 146 | const auto offset = float(oq.offset) - scale0 * float(iq0.offset) - scale1 * float(iq1.offset); |
| 147 | const auto max_acc = (std::abs(scale0) + std::abs(scale1)) * 1024.f + std::abs(offset); |
| 148 | |
| 149 | if(max_acc > 2097151.f) // 2^21 - 1 |
| 150 | { |
| 151 | // It might not be possible to store the result as 22.10 signed fixed-point number. |
| 152 | return false; |
| 153 | } |
| 154 | |
| 155 | return true; |
| 156 | } |
| 157 | |
| 158 | template <typename ScalarType> |
| 159 | void add_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) |
| 160 | { |
| 161 | ARM_COMPUTE_UNUSED(policy); |
| 162 | |
| 163 | const auto in0_info = src0->info(); |
| 164 | const auto in1_info = src1->info(); |
| 165 | |
| 166 | const auto &in0_shape = in0_info->tensor_shape(); |
| 167 | const auto &in1_shape = in1_info->tensor_shape(); |
| 168 | |
| 169 | // Create input windows. |
| 170 | Window in0_win = window.broadcast_if_dimension_le_one(in0_shape); |
| 171 | Window in1_win = window.broadcast_if_dimension_le_one(in1_shape); |
| 172 | |
| 173 | // Clear the x dimension on the execution window as we process the whole row each iteration. |
| 174 | Window win = window; |
| 175 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 176 | |
| 177 | constexpr int window_step_x = 16; |
| 178 | const auto window_start_x = window.x().start(); |
| 179 | const auto window_end_x = window.x().end(); |
| 180 | const auto is_broadcast_across_x = in0_shape.x() != in1_shape.x(); |
| 181 | |
| 182 | const auto iq0_info = in0_info->quantization_info().uniform(); |
| 183 | const auto iq1_info = in1_info->quantization_info().uniform(); |
| 184 | const auto oq_info = dst->info()->quantization_info().uniform(); |
| 185 | |
| 186 | const auto in0_scale = iq0_info.scale / oq_info.scale; |
| 187 | const auto in1_scale = iq1_info.scale / oq_info.scale; |
| 188 | const auto offset = float(oq_info.offset) - in0_scale * float(iq0_info.offset) - in1_scale * float(iq1_info.offset); |
| 189 | |
| 190 | const auto in0_scale_6p10 = static_cast<int16_t>(support::cpp11::lround(in0_scale * 1024.f)); |
| 191 | const auto in1_scale_6p10 = static_cast<int16_t>(support::cpp11::lround(in1_scale * 1024.f)); |
| 192 | const auto offset_22p10 = static_cast<int32_t>(support::cpp11::lround(offset * 1024.f)); |
| 193 | |
| 194 | if(is_broadcast_across_x) |
| 195 | { |
| 196 | // Prefix: a = non-broadcast, b = broadcast. |
| 197 | |
| 198 | const auto is_broadcast_input_1 = in1_win.x().step() == 0; |
| 199 | auto a_win = is_broadcast_input_1 ? in0_win : in1_win; |
| 200 | auto b_win = is_broadcast_input_1 ? in1_win : in0_win; |
| 201 | const auto a_tensor = is_broadcast_input_1 ? src0 : src1; |
| 202 | const auto b_tensor = is_broadcast_input_1 ? src1 : src0; |
| 203 | |
| 204 | const auto a_scale_6p10 = is_broadcast_input_1 ? in0_scale_6p10 : in1_scale_6p10; |
| 205 | const auto b_scale = is_broadcast_input_1 ? in1_scale : in0_scale; |
| 206 | const auto a_vscale_6p10 = wrapper::vdup_n(a_scale_6p10, wrapper::traits::vector_64_tag()); |
| 207 | |
Viet-Hoa Do | 910e3f9 | 2022-10-11 13:21:35 +0100 | [diff] [blame^] | 208 | #ifndef __aarch64__ |
| 209 | const auto a_scale = is_broadcast_input_1 ? in0_scale : in1_scale; |
| 210 | #endif // __aarch64__ |
| 211 | |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 212 | // Clear the x dimension on the execution window as we process the whole row each iteration. |
| 213 | a_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 214 | |
| 215 | Iterator a_input_it(a_tensor, a_win); |
| 216 | Iterator b_input_it(b_tensor, b_win); |
| 217 | Iterator out_it(dst, win); |
| 218 | |
| 219 | execute_window_loop(win, [&](const Coordinates &) |
| 220 | { |
| 221 | const auto a_ptr = reinterpret_cast<const ScalarType *>(a_input_it.ptr()); |
| 222 | const auto b_ptr = reinterpret_cast<const ScalarType *>(b_input_it.ptr()); |
| 223 | const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()); |
| 224 | |
| 225 | const auto b_val = *b_ptr; |
Viet-Hoa Do | 910e3f9 | 2022-10-11 13:21:35 +0100 | [diff] [blame^] | 226 | const auto b_scaled = b_scale * b_val; |
| 227 | const auto b_scaled_22p10 = static_cast<int32_t>(support::cpp11::lround(b_scaled * 1024.f)); |
| 228 | const auto b_scaled_offseted_22p10 = b_scaled_22p10 + offset_22p10; |
| 229 | const auto b_vscaled_offseted_22p10 = wrapper::vdup_n(b_scaled_offseted_22p10, wrapper::traits::vector_128_tag()); |
| 230 | |
| 231 | #ifndef __aarch64__ |
| 232 | const auto b_scaled_offseted = b_scaled + offset; |
| 233 | #endif // __aarch64__ |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 234 | |
| 235 | int x = window_start_x; |
| 236 | |
| 237 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 238 | { |
| 239 | // Load the input. |
| 240 | const auto a_vin_8p0 = wrapper::vloadq(a_ptr + x); |
| 241 | |
| 242 | // Widen the non-broadcast elements to signed 16-bit regardless of the input signedness. |
| 243 | const auto a_vin_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(a_vin_8p0))); |
| 244 | const auto a_vin_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(a_vin_8p0))); |
| 245 | |
| 246 | // Multiply the non-broadcast elements by the scale factor, add the scaled broadcast elements and the offset. |
| 247 | // Widen and store the result in 32-bit integer. |
| 248 | const auto vout_22p10_00 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgetlow(a_vin_16p0_0), a_vscale_6p10); |
| 249 | const auto vout_22p10_01 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgethigh(a_vin_16p0_0), a_vscale_6p10); |
| 250 | const auto vout_22p10_10 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgetlow(a_vin_16p0_1), a_vscale_6p10); |
| 251 | const auto vout_22p10_11 = wrapper::vmlal(b_vscaled_offseted_22p10, wrapper::vgethigh(a_vin_16p0_1), a_vscale_6p10); |
| 252 | |
| 253 | // Remove 2 bits of the fractional part, round, narrow to 16-bit and saturate the result. |
| 254 | const auto vout_8p8_0 = wrapper::vcombine( |
| 255 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_00), |
| 256 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_01) |
| 257 | ); |
| 258 | const auto vout_8p8_1 = wrapper::vcombine( |
| 259 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_10), |
| 260 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_11) |
| 261 | ); |
| 262 | |
| 263 | // Remove 8 bits of the fractional part, round, narrow to 8-bit and saturate the result. |
| 264 | const auto vout_8p0 = wrapper::vcombine( |
| 265 | wrapper::vqrshrn<8>(vout_8p8_0), |
| 266 | wrapper::vqrshrn<8>(vout_8p8_1) |
| 267 | ); |
| 268 | |
| 269 | // Store the result. |
| 270 | wrapper::vstore(out_ptr + x, vout_8p0); |
| 271 | } |
| 272 | |
| 273 | // Process the left-over elements. |
| 274 | for(; x < window_end_x; ++x) |
| 275 | { |
Viet-Hoa Do | 910e3f9 | 2022-10-11 13:21:35 +0100 | [diff] [blame^] | 276 | #ifdef __aarch64__ |
| 277 | out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<2, ScalarType>(int32_t(a_ptr[x]) * a_scale_6p10 + b_scaled_offseted_22p10)); |
| 278 | #else // __aarch64__ |
| 279 | out_ptr[x] = utility::clamp<int, ScalarType>(support::cpp11::lround(float(a_ptr[x]) * a_scale + b_scaled_offseted)); |
| 280 | #endif // __aarch64__ |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 281 | } |
| 282 | }, |
| 283 | b_input_it, a_input_it, out_it); |
| 284 | } |
| 285 | else |
| 286 | { |
| 287 | const auto vscale0_6p10 = wrapper::vdup_n(in0_scale_6p10, wrapper::traits::vector_64_tag()); |
| 288 | const auto vscale1_6p10 = wrapper::vdup_n(in1_scale_6p10, wrapper::traits::vector_64_tag()); |
| 289 | const auto voffset_22p10 = wrapper::vdup_n(offset_22p10, wrapper::traits::vector_128_tag()); |
| 290 | |
| 291 | // Clear the x dimension on the execution window as we process the whole row each iteration. |
| 292 | in0_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 293 | in1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 294 | |
| 295 | Iterator in0_it(src0, in0_win); |
| 296 | Iterator in1_it(src1, in1_win); |
| 297 | Iterator out_it(dst, win); |
| 298 | |
| 299 | execute_window_loop(win, [&](const Coordinates &) |
| 300 | { |
| 301 | const auto in0_ptr = reinterpret_cast<const ScalarType *>(in0_it.ptr()); |
| 302 | const auto in1_ptr = reinterpret_cast<const ScalarType *>(in1_it.ptr()); |
| 303 | const auto out_ptr = reinterpret_cast<ScalarType *>(out_it.ptr()); |
| 304 | |
| 305 | int x = window_start_x; |
| 306 | |
| 307 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 308 | { |
| 309 | // Load the inputs. |
| 310 | const auto vin0_8p0 = wrapper::vloadq(in0_ptr + x); |
| 311 | const auto vin1_8p0 = wrapper::vloadq(in1_ptr + x); |
| 312 | |
| 313 | // Widen the input elements to signed 16-bit regardless of the input signedness. |
| 314 | const auto vin0_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin0_8p0))); |
| 315 | const auto vin0_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin0_8p0))); |
| 316 | const auto vin1_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin1_8p0))); |
| 317 | const auto vin1_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin1_8p0))); |
| 318 | |
| 319 | // Multiply the input elements by the scale factor and add the offset. |
| 320 | // Widen and store the result in 32-bit integer. |
| 321 | const auto vscaled0_offseted_22p10_00 = wrapper::vmlal(voffset_22p10, wrapper::vgetlow(vin0_16p0_0), vscale0_6p10); |
| 322 | const auto vscaled0_offseted_22p10_01 = wrapper::vmlal(voffset_22p10, wrapper::vgethigh(vin0_16p0_0), vscale0_6p10); |
| 323 | const auto vscaled0_offseted_22p10_10 = wrapper::vmlal(voffset_22p10, wrapper::vgetlow(vin0_16p0_1), vscale0_6p10); |
| 324 | const auto vscaled0_offseted_22p10_11 = wrapper::vmlal(voffset_22p10, wrapper::vgethigh(vin0_16p0_1), vscale0_6p10); |
| 325 | |
| 326 | const auto vout_22p10_00 = wrapper::vmlal(vscaled0_offseted_22p10_00, wrapper::vgetlow(vin1_16p0_0), vscale1_6p10); |
| 327 | const auto vout_22p10_01 = wrapper::vmlal(vscaled0_offseted_22p10_01, wrapper::vgethigh(vin1_16p0_0), vscale1_6p10); |
| 328 | const auto vout_22p10_10 = wrapper::vmlal(vscaled0_offseted_22p10_10, wrapper::vgetlow(vin1_16p0_1), vscale1_6p10); |
| 329 | const auto vout_22p10_11 = wrapper::vmlal(vscaled0_offseted_22p10_11, wrapper::vgethigh(vin1_16p0_1), vscale1_6p10); |
| 330 | |
| 331 | // Remove 2 bits of the fractional part, round, narrow to 16-bit and saturate the result. |
| 332 | const auto vout_8p8_0 = wrapper::vcombine( |
| 333 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_00), |
| 334 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_01) |
| 335 | ); |
| 336 | const auto vout_8p8_1 = wrapper::vcombine( |
| 337 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_10), |
| 338 | wrapper::vqrshrn_ex<2, ScalarType>(vout_22p10_11) |
| 339 | ); |
| 340 | |
| 341 | // Remove 8 bits of the fractional part, round, narrow to 8-bit and saturate the result. |
| 342 | const auto vout_8p0 = wrapper::vcombine( |
| 343 | wrapper::vqrshrn<8>(vout_8p8_0), |
| 344 | wrapper::vqrshrn<8>(vout_8p8_1) |
| 345 | ); |
| 346 | |
| 347 | // Store the result. |
| 348 | wrapper::vstore(out_ptr + x, vout_8p0); |
| 349 | } |
| 350 | |
| 351 | // Process the left-over elements. |
| 352 | for(; x < window_end_x; ++x) |
| 353 | { |
Viet-Hoa Do | 910e3f9 | 2022-10-11 13:21:35 +0100 | [diff] [blame^] | 354 | #ifdef __aarch64__ |
| 355 | out_ptr[x] = wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<2, ScalarType>(int32_t(in0_ptr[x]) * in0_scale_6p10 + int32_t(in1_ptr[x]) * in1_scale_6p10 + offset_22p10)); |
| 356 | #else // __aarch64__ |
| 357 | out_ptr[x] = utility::clamp<int, ScalarType>(support::cpp11::lround(float(in0_ptr[x]) * in0_scale + float(in1_ptr[x]) * in1_scale + offset)); |
| 358 | #endif // __aarch64__ |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 359 | } |
| 360 | }, |
| 361 | in0_it, in1_it, out_it); |
| 362 | } |
| 363 | } |
| 364 | |
Dana Zlotnik | bd2942d | 2021-11-15 08:46:04 +0200 | [diff] [blame] | 365 | template void add_same_neon<float>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 366 | template void add_same_neon<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 367 | template void add_same_neon<int32_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 368 | template void add_same_neon<int16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 369 | |
| 370 | #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) |
| 371 | template void add_same_neon<float16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 372 | #endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ |
| 373 | |
Viet-Hoa Do | 40b4419 | 2022-09-22 10:24:23 +0100 | [diff] [blame] | 374 | template void add_q8_neon_fixedpoint<int8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 375 | template void add_q8_neon_fixedpoint<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window); |
| 376 | |
Michalis Spyrou | a3c9a3b | 2020-12-08 21:02:16 +0000 | [diff] [blame] | 377 | } // namespace cpu |
| 378 | } // namespace arm_compute |