Hu Chunming / vpt_ascend

functors.hpp 8.45 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

#ifndef OPENCV_DNN_SRC_CUDA_FUNCTORS_HPP
#define OPENCV_DNN_SRC_CUDA_FUNCTORS_HPP

#include <cuda_runtime.h>

#include "math.hpp"

#include "../cuda4dnn/csl/nvcc_defs.hpp"

namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {

template <class T>
struct IdentityFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE IdentityFunctor() { }
    CUDA4DNN_DEVICE IdentityFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        return value;
    };
};

template <class T>
struct ReLUFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() : slope(0) { }
        CUDA4DNN_HOST_DEVICE Params(T slope_) : slope(slope_) { }
        T slope;
    };

    CUDA4DNN_DEVICE ReLUFunctor() : ReLUFunctor(Params{}) { }
    CUDA4DNN_DEVICE ReLUFunctor(const Params& params) : slope(params.slope) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::log1pexp;
        return value >= T(0) ? value : slope * value;
    }

    T slope;
};

template <class T>
struct ClippedReLUFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() : floor(0), ceiling(6) { }
        CUDA4DNN_HOST_DEVICE Params(T floor_, T ceiling_) : floor(floor_), ceiling(ceiling_) { }
        T floor, ceiling;
    };

    CUDA4DNN_DEVICE ClippedReLUFunctor() : ClippedReLUFunctor(Params{}) { }
    CUDA4DNN_DEVICE ClippedReLUFunctor(const Params& params) : floor{params.floor}, ceiling{params.ceiling} { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::clamp;
        return clamp(value, floor, ceiling);
    }

    T floor, ceiling;
};

template <class T>
struct TanHFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE TanHFunctor() { }
    CUDA4DNN_DEVICE TanHFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::tanh;
        return tanh(value);
    }
};

template <class T>
struct SwishFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE SwishFunctor() { }
    CUDA4DNN_DEVICE SwishFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        // f(x) = x * sigmoid(x)
        using csl::device::fast_divide;
        using csl::device::fast_exp;
        return fast_divide(value, static_cast<T>(1) + fast_exp(-value));
    }
};

template <class T>
struct MishFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE MishFunctor() { }
    CUDA4DNN_DEVICE MishFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::tanh;
        using csl::device::log1pexp;
        return value * tanh(log1pexp(value));
    }
};

template <>
struct MishFunctor<float> {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE MishFunctor() { }
    CUDA4DNN_DEVICE MishFunctor(const Params& params) { }

    CUDA4DNN_DEVICE float operator()(float value) {
        // f(x) = x * tanh(log1pexp(x));
        using csl::device::fast_divide;
        using csl::device::fast_exp;

        auto e = fast_exp(value);
        auto n = e * e + 2 * e;
        if (value <= -0.6f)
            return value * fast_divide(n, n + 2);
        return value - 2 * fast_divide(value, n + 2);
    }
};

#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
template <>
struct MishFunctor<__half> {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE MishFunctor() { }
    CUDA4DNN_DEVICE MishFunctor(const Params& params) { }

    CUDA4DNN_DEVICE __half operator()(__half value) {
        return MishFunctor<float>()(value);
    }
};
#endif

template <class T>
struct SigmoidFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE SigmoidFunctor() { }
    CUDA4DNN_DEVICE SigmoidFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::fast_sigmoid;
        return fast_sigmoid(value);
    }
};

template <class T>
struct ELUFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE ELUFunctor() { }
    CUDA4DNN_DEVICE ELUFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::expm1;
        return value >= T(0) ? value : expm1(value);
    }
};

template <class T>
struct AbsFunctor {
    struct Params { };

    CUDA4DNN_DEVICE AbsFunctor() { }
    CUDA4DNN_DEVICE AbsFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::abs;
        return abs(value);
    }
};

template <class T>
struct BNLLFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE BNLLFunctor() { }
    CUDA4DNN_DEVICE BNLLFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::log1pexp;
        return value > T(0) ? value + log1pexp(-value) : log1pexp(value);
    }
};

template <class T>
struct PowerFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() : exp(1), scale(1), shift(0) { }
        CUDA4DNN_HOST_DEVICE Params(T exp_, T scale_, T shift_) : exp(exp_), scale(scale_), shift(shift_) { }
        T exp, scale, shift;
    };

    CUDA4DNN_DEVICE PowerFunctor() : PowerFunctor(Params{}) { }
    CUDA4DNN_DEVICE PowerFunctor(const Params& params) : exp{params.exp}, scale{params.scale}, shift{params.shift} { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::pow;
        return pow(shift + scale * value, exp);
    }

    T exp, scale, shift;
};

template <class T>
struct ExpFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() : normScale(1), normShift(0) { }
        CUDA4DNN_HOST_DEVICE Params(T nScale_, T nShift_) : normScale(nScale_), normShift(nShift_) { }
        T normScale, normShift;
    };

    CUDA4DNN_DEVICE ExpFunctor() : ExpFunctor(Params{}) { }
    CUDA4DNN_DEVICE ExpFunctor(const Params& params) : normScale{params.normScale}, normShift{params.normShift} { }

    CUDA4DNN_DEVICE T operator()(T value) {
        using csl::device::fast_exp;
        return fast_exp(normShift + normScale * value);
    }

    T normScale, normShift;
};

template <class T>
struct MaxFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE MaxFunctor() { }
    CUDA4DNN_DEVICE MaxFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T x, T y) {
        using csl::device::max;
        return max(x, y);
    }
};

template <class T>
struct MinFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE MinFunctor() { }
    CUDA4DNN_DEVICE MinFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T x, T y) {
        using csl::device::min;
        return min(x, y);
    }
};

template <class T>
struct SumFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE SumFunctor() { }
    CUDA4DNN_DEVICE SumFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T x, T y) { return x + y; }
};

template <class T>
struct ScaledSumFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() : scale_x(1), scale_y(1) { }
        CUDA4DNN_HOST_DEVICE Params(T scale_x_, T scale_y_) : scale_x(scale_x_), scale_y(scale_y_) { }
        T scale_x, scale_y;
    };

    CUDA4DNN_DEVICE ScaledSumFunctor() : scale_x(1), scale_y(1) { }
    CUDA4DNN_DEVICE ScaledSumFunctor(const Params& params) : scale_x{params.scale_x}, scale_y{params.scale_y} { }

    CUDA4DNN_DEVICE T operator()(T x, T y) { return scale_x * x + scale_y * y; }

    T scale_x, scale_y;
};

template <class T>
struct ProductFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE ProductFunctor() { }
    CUDA4DNN_DEVICE ProductFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T x, T y) { return x * y; }
};

template <class T>
struct DivFunctor {
    struct Params {
        CUDA4DNN_HOST_DEVICE Params() { }
    };

    CUDA4DNN_DEVICE DivFunctor() { }
    CUDA4DNN_DEVICE DivFunctor(const Params& params) { }

    CUDA4DNN_DEVICE T operator()(T x, T y) { return x / y; }
};

}}}} /* namespace cv::dnn::cuda4dnn::kernels */

#endif /* OPENCV_DNN_SRC_CUDA_FUNCTORS_HPP */