Michalis Spyrou | c4d4555 | 2020-10-19 12:41:30 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2020 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 | #include "arm_compute/core/Helpers.h" |
| 25 | #include "arm_compute/core/ITensorPack.h" |
| 26 | #include "arm_compute/core/Window.h" |
| 27 | #include "src/core/NEON/NEMath.h" |
| 28 | #include "src/core/NEON/wrapper/wrapper.h" |
| 29 | #include "src/core/common/StdTypes.h" |
| 30 | #include "src/core/common/Validate.h" |
| 31 | |
| 32 | #include <arm_neon.h> |
| 33 | #include <cmath> |
| 34 | #include <cstddef> |
| 35 | |
| 36 | namespace arm_compute |
| 37 | { |
| 38 | namespace cpu |
| 39 | { |
| 40 | namespace |
| 41 | { |
| 42 | #ifndef __aarch64__ |
| 43 | inline float32x4_t mask_float_vector(const float32x4_t &in, const uint32x4_t &mask) |
| 44 | { |
| 45 | auto int_in = vreinterpretq_u32_f32(in); |
| 46 | return vreinterpretq_f32_u32(wrapper::vand(int_in, mask)); |
| 47 | } |
| 48 | #endif /* __arch64__ */ |
| 49 | } // namespace |
| 50 | |
| 51 | void fp32_neon_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) |
| 52 | { |
| 53 | /** NEON vector tag type. */ |
| 54 | using ExactTagType = typename arm_compute::wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>; |
| 55 | |
| 56 | constexpr int window_step_x = 4; |
| 57 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 58 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 59 | const ActivationLayerInfo::ActivationFunction act = act_info.activation(); |
| 60 | |
| 61 | Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); |
| 62 | win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 63 | |
| 64 | Iterator input(src, win_collapsed); |
| 65 | Iterator output(dst, win_collapsed); |
| 66 | |
| 67 | // In case of non-aarch64, a small delta value is added to the input |
| 68 | // to prevent NAN values caused by zeros in inputs to SQRT. |
| 69 | // In case of aarh64, we call vsqrt directly, so we don't use delta. |
| 70 | #ifndef __aarch64__ |
| 71 | const auto delta = wrapper::vdup_n(static_cast<float>(1e-24), ExactTagType {}); |
| 72 | #endif /* __aarch64 */ |
| 73 | const auto const_1 = wrapper::vdup_n(static_cast<float>(1.f), ExactTagType {}); |
| 74 | const auto const_0 = wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{}); |
| 75 | const auto const_6 = wrapper::vdup_n(static_cast<float>(6.f), ExactTagType{}); |
| 76 | const auto const_3 = wrapper::vdup_n(static_cast<float>(3.f), ExactTagType{}); |
| 77 | const auto const_inv_6 = wrapper::vdup_n(static_cast<float>(0.166666667f), ExactTagType{}); |
| 78 | |
| 79 | const auto va = wrapper::vdup_n(static_cast<float>(act_info.a()), ExactTagType{}); |
| 80 | const auto vb = wrapper::vdup_n(static_cast<float>(act_info.b()), ExactTagType{}); |
| 81 | const auto a = static_cast<float>(act_info.a()); |
| 82 | const auto b = static_cast<float>(act_info.b()); |
| 83 | execute_window_loop(win_collapsed, [&](const Coordinates &) |
| 84 | { |
| 85 | const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); |
| 86 | const auto output_ptr = reinterpret_cast<float *>(output.ptr()); |
| 87 | |
| 88 | wrapper::traits::neon_bitvector_t<float, wrapper::traits::BitWidth::W128> tmp; |
| 89 | |
| 90 | // Compute S elements per iteration |
| 91 | int x = window_start_x; |
| 92 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 93 | { |
| 94 | const auto vin = wrapper::vloadq(input_ptr + x); |
| 95 | switch(act) |
| 96 | { |
| 97 | case ActivationLayerInfo::ActivationFunction::ABS: |
| 98 | tmp = wrapper::vabs(vin); |
| 99 | break; |
| 100 | case ActivationLayerInfo::ActivationFunction::LINEAR: |
| 101 | tmp = wrapper::vmla(vb, va, vin); |
| 102 | break; |
| 103 | case ActivationLayerInfo::ActivationFunction::LOGISTIC: |
| 104 | tmp = wrapper::vinv(wrapper::vadd(const_1, wrapper::vexpq(wrapper::vneg(vin)))); |
| 105 | break; |
| 106 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 107 | tmp = wrapper::vmax(const_0, vin); |
| 108 | break; |
| 109 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 110 | tmp = wrapper::vmin(va, wrapper::vmax(const_0, vin)); |
| 111 | break; |
| 112 | case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: |
| 113 | tmp = wrapper::vmin(va, wrapper::vmax(vb, vin)); |
| 114 | break; |
| 115 | case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: |
| 116 | tmp = wrapper::vbsl(wrapper::vcgt(vin, const_0), vin, wrapper::vmul(va, vin)); |
| 117 | break; |
| 118 | case ActivationLayerInfo::ActivationFunction::SOFT_RELU: |
| 119 | tmp = wrapper::vlog(wrapper::vadd(const_1, wrapper::vexpq(vin))); |
| 120 | break; |
| 121 | case ActivationLayerInfo::ActivationFunction::ELU: |
| 122 | tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); |
| 123 | break; |
| 124 | case ActivationLayerInfo::ActivationFunction::SQRT: |
| 125 | #ifdef __aarch64__ |
| 126 | tmp = wrapper::vsqrt(vin); |
| 127 | #else /* aarch64 */ |
| 128 | { |
| 129 | const auto bitmask = wrapper::vceq(vin, wrapper::vdup_n(0.f, ExactTagType{})); |
| 130 | tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, mask_float_vector(delta, bitmask)))); |
| 131 | tmp = mask_float_vector(tmp, wrapper::vnot(bitmask)); |
| 132 | } |
| 133 | #endif /* aarch64 */ |
| 134 | break; |
| 135 | case ActivationLayerInfo::ActivationFunction::SQUARE: |
| 136 | tmp = wrapper::vmul(vin, vin); |
| 137 | break; |
| 138 | case ActivationLayerInfo::ActivationFunction::TANH: |
| 139 | tmp = wrapper::vmul(va, wrapper::vtanh(wrapper::vmul(vb, vin))); |
| 140 | break; |
| 141 | case ActivationLayerInfo::ActivationFunction::IDENTITY: |
| 142 | tmp = vin; |
| 143 | break; |
| 144 | case ActivationLayerInfo::ActivationFunction::HARD_SWISH: |
| 145 | tmp = wrapper::vmul(vin, wrapper::vmul(const_inv_6, wrapper::vmin(const_6, wrapper::vmax(const_0, wrapper::vadd(vin, const_3))))); |
| 146 | break; |
| 147 | default: |
| 148 | ARM_COMPUTE_ERROR("Unsupported activation function"); |
| 149 | } |
| 150 | wrapper::vstore(output_ptr + x, tmp); |
| 151 | } |
| 152 | |
| 153 | // Compute left-over elements |
| 154 | for(; x < window_end_x; ++x) |
| 155 | { |
| 156 | const float in = *(reinterpret_cast<const float *>(input_ptr + x)); |
| 157 | float tmp; |
| 158 | switch(act) |
| 159 | { |
| 160 | case ActivationLayerInfo::ActivationFunction::ABS: |
| 161 | tmp = std::abs(in); |
| 162 | break; |
| 163 | case ActivationLayerInfo::ActivationFunction::LINEAR: |
| 164 | tmp = a * in + b; |
| 165 | break; |
| 166 | case ActivationLayerInfo::ActivationFunction::LOGISTIC: |
| 167 | tmp = static_cast<float>(1) / (static_cast<float>(1) + std::exp(-in)); |
| 168 | break; |
| 169 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 170 | tmp = std::max<float>(static_cast<float>(0), in); |
| 171 | break; |
| 172 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 173 | tmp = std::min<float>(a, std::max(static_cast<float>(0), in)); |
| 174 | break; |
| 175 | case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: |
| 176 | tmp = std::min<float>(a, std::max<float>(b, in)); |
| 177 | break; |
| 178 | case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: |
| 179 | tmp = (in > 0) ? in : a * in; |
| 180 | break; |
| 181 | case ActivationLayerInfo::ActivationFunction::SOFT_RELU: |
| 182 | tmp = std::log(static_cast<float>(1) + std::exp(in)); |
| 183 | break; |
| 184 | case ActivationLayerInfo::ActivationFunction::ELU: |
| 185 | tmp = (in >= 0) ? in : a * (std::exp(in) - 1); |
| 186 | break; |
| 187 | case ActivationLayerInfo::ActivationFunction::SQRT: |
| 188 | tmp = std::sqrt(in); |
| 189 | break; |
| 190 | case ActivationLayerInfo::ActivationFunction::SQUARE: |
| 191 | tmp = in * in; |
| 192 | break; |
| 193 | case ActivationLayerInfo::ActivationFunction::TANH: |
| 194 | tmp = a * std::tanh(b * in); |
| 195 | break; |
| 196 | case ActivationLayerInfo::ActivationFunction::IDENTITY: |
| 197 | tmp = in; |
| 198 | break; |
| 199 | case ActivationLayerInfo::ActivationFunction::HARD_SWISH: |
| 200 | tmp = in * ((std::min(std::max((in + 3), 0.0f), 6.0f)) * 0.166666667f); |
| 201 | break; |
| 202 | default: |
| 203 | ARM_COMPUTE_ERROR("Unsupported activation function"); |
| 204 | } |
| 205 | *(output_ptr + x) = tmp; |
| 206 | } |
| 207 | }, |
| 208 | input, output); |
| 209 | } |
| 210 | } // namespace cpu |
| 211 | } // namespace arm_compute |