| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *ocl_eltwise_header = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2020-2022 Intel Corporation )==" "\n" |
| 7 | R"==(* )==" "\n" |
| 8 | R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==" "\n" |
| 9 | R"==(* you may not use this file except in compliance with the License. )==" "\n" |
| 10 | R"==(* You may obtain a copy of the License at )==" "\n" |
| 11 | R"==(* )==" "\n" |
| 12 | R"==(* http: )==" "\n" |
| 13 | R"==(* )==" "\n" |
| 14 | R"==(* Unless required by applicable law or agreed to in writing, software )==" "\n" |
| 15 | R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==" "\n" |
| 16 | R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==" "\n" |
| 17 | R"==(* See the License for the specific language governing permissions and )==" "\n" |
| 18 | R"==(* limitations under the License. )==" "\n" |
| 19 | R"==(*******************************************************************************/ )==" "\n" |
| 20 | R"==(#ifndef GPU_OCL_OCL_ELTWISE_H )==" "\n" |
| 21 | R"==(#define GPU_OCL_OCL_ELTWISE_H )==" "\n" |
| 22 | R"==(#if WITH_ELTWISE )==" "\n" |
| 23 | R"==(#if DT_F16 == 1 )==" "\n" |
| 24 | R"==(#pragma OPENCL EXTENSION cl_khr_fp16 : enable )==" "\n" |
| 25 | R"==(#endif )==" "\n" |
| 26 | R"==(#ifndef DATA_MAX )==" "\n" |
| 27 | R"==(#if DT_F16 == 1 )==" "\n" |
| 28 | R"==(#define DATA_MAX HALF_MAX )==" "\n" |
| 29 | R"==(#elif DT_S8 == 1 )==" "\n" |
| 30 | R"==(#define DATA_MAX CHAR_MAX )==" "\n" |
| 31 | R"==(#elif DT_U8 == 1 )==" "\n" |
| 32 | R"==(#define DATA_MAX UCHAR_MAX )==" "\n" |
| 33 | R"==(#else )==" "\n" |
| 34 | R"==(#define DATA_MAX FLT_MAX )==" "\n" |
| 35 | R"==(#endif )==" "\n" |
| 36 | R"==(#endif )==" "\n" |
| 37 | R"==(float relu_fwd(float s, float alpha) { )==" "\n" |
| 38 | R"==(return s > 0 ? s : ((alpha == 0) ? 0 : s * alpha); )==" "\n" |
| 39 | R"==(} )==" "\n" |
| 40 | R"==(float relu_bwd(float dd, float s, float alpha) { )==" "\n" |
| 41 | R"==(return s > 0 ? dd : dd * alpha; )==" "\n" |
| 42 | R"==(} )==" "\n" |
| 43 | R"==(float relu_bwd_use_dst(float dd, float d, float alpha) { )==" "\n" |
| 44 | R"==(return d > 0 ? dd : dd * alpha; )==" "\n" |
| 45 | R"==(} )==" "\n" |
| 46 | R"==(float linear_fwd(float s, float alpha, float beta) { )==" "\n" |
| 47 | R"==(return alpha * s + beta; )==" "\n" |
| 48 | R"==(} )==" "\n" |
| 49 | R"==(float linear_bwd(float dd, float alpha) { )==" "\n" |
| 50 | R"==(return dd * alpha; )==" "\n" |
| 51 | R"==(} )==" "\n" |
| 52 | R"==(float soft_relu_fwd(float s, float alpha) { )==" "\n" |
| 53 | R"==(s = alpha * s; )==" "\n" |
| 54 | R"==(float v = (s < log((float)DATA_MAX) ? log1p(exp(s)) : s); )==" "\n" |
| 55 | R"==(return v / alpha; )==" "\n" |
| 56 | R"==(} )==" "\n" |
| 57 | R"==(float soft_relu_bwd(float dd, float s, float alpha) { )==" "\n" |
| 58 | R"==(s = alpha * s; )==" "\n" |
| 59 | R"==(return dd / (1 + exp(-s)); )==" "\n" |
| 60 | R"==(} )==" "\n" |
| 61 | R"==(float logistic_fwd(float s) { )==" "\n" |
| 62 | R"==(return 1.0f / (1.0f + exp(-s)); )==" "\n" |
| 63 | R"==(} )==" "\n" |
| 64 | R"==(float logistic_bwd(float dd, float s) { )==" "\n" |
| 65 | R"==(float v = logistic_fwd(s); )==" "\n" |
| 66 | R"==(return dd * v * (1 - v); )==" "\n" |
| 67 | R"==(} )==" "\n" |
| 68 | R"==(float logistic_bwd_use_dst(float dd, float d) { )==" "\n" |
| 69 | R"==(return dd * d * (1 - d); )==" "\n" |
| 70 | R"==(} )==" "\n" |
| 71 | R"==(float square_fwd(float s) { )==" "\n" |
| 72 | R"==(return s * s; )==" "\n" |
| 73 | R"==(} )==" "\n" |
| 74 | R"==(float square_bwd(float dd, float s) { )==" "\n" |
| 75 | R"==(return dd * 2 * s; )==" "\n" |
| 76 | R"==(} )==" "\n" |
| 77 | R"==(float sqrt_fwd(float s) { )==" "\n" |
| 78 | R"==(return sqrt(s); )==" "\n" |
| 79 | R"==(} )==" "\n" |
| 80 | R"==(float sqrt_bwd(float dd, float s) { )==" "\n" |
| 81 | R"==(return dd / (2 * sqrt(s)); )==" "\n" |
| 82 | R"==(} )==" "\n" |
| 83 | R"==(float sqrt_bwd_use_dst(float dd, float d) { )==" "\n" |
| 84 | R"==(return dd / (2 * d); )==" "\n" |
| 85 | R"==(} )==" "\n" |
| 86 | R"==(float abs_fwd(float s) { )==" "\n" |
| 87 | R"==(return s > 0 ? s : -s; )==" "\n" |
| 88 | R"==(} )==" "\n" |
| 89 | R"==(float abs_bwd(float dd, float s) { )==" "\n" |
| 90 | R"==(return s > 0 ? dd : s < 0 ? -dd : 0; )==" "\n" |
| 91 | R"==(} )==" "\n" |
| 92 | R"==(float tanh_fwd(float s) { )==" "\n" |
| 93 | R"==(return tanh(s); )==" "\n" |
| 94 | R"==(} )==" "\n" |
| 95 | R"==(float tanh_bwd(float dd, float s) { )==" "\n" |
| 96 | R"==(float e = tanh_fwd(s); )==" "\n" |
| 97 | R"==(return dd * (1 - e) * (1 + e); )==" "\n" |
| 98 | R"==(} )==" "\n" |
| 99 | R"==(float tanh_bwd_use_dst(float dd, float d) { )==" "\n" |
| 100 | R"==(return dd * (1 - d) * (1 + d); )==" "\n" |
| 101 | R"==(} )==" "\n" |
| 102 | R"==(float mish_fwd(float s) { )==" "\n" |
| 103 | R"==(return s * tanh_fwd(soft_relu_fwd(s, 1.f)); )==" "\n" |
| 104 | R"==(} )==" "\n" |
| 105 | R"==(float mish_bwd(float dd, float s) { )==" "\n" |
| 106 | R"==(const float tanh = tanh_fwd(soft_relu_fwd(s, 1.f)); )==" "\n" |
| 107 | R"==(const float srelu_bwd = soft_relu_bwd(1.f, s, 1.f); )==" "\n" |
| 108 | R"==(const float derivative = tanh + s * srelu_bwd * (1 - pow(tanh, 2.0f)); )==" "\n" |
| 109 | R"==(return dd * derivative; )==" "\n" |
| 110 | R"==(} )==" "\n" |
| 111 | R"==(float elu_fwd(float s, float alpha) { )==" "\n" |
| 112 | R"==(return s > 0 ? s : alpha * expm1(s); )==" "\n" |
| 113 | R"==(} )==" "\n" |
| 114 | R"==(float elu_bwd(float dd, float s, float alpha) { )==" "\n" |
| 115 | R"==(return dd * (s > 0 ? 1 : alpha * exp(s)); )==" "\n" |
| 116 | R"==(} )==" "\n" |
| 117 | R"==(float elu_bwd_use_dst(float dd, float d, float alpha) { )==" "\n" |
| 118 | R"==(return dd * (d > 0 ? 1 : d + alpha); )==" "\n" |
| 119 | R"==(} )==" "\n" |
| 120 | R"==(float exp_fwd(float s) { )==" "\n" |
| 121 | R"==(return exp(s); )==" "\n" |
| 122 | R"==(} )==" "\n" |
| 123 | R"==(float exp_bwd(float dd, float s) { )==" "\n" |
| 124 | R"==(return dd * exp_fwd(s); )==" "\n" |
| 125 | R"==(} )==" "\n" |
| 126 | R"==(float exp_bwd_use_dst(float dd, float d) { )==" "\n" |
| 127 | R"==(return dd * d; )==" "\n" |
| 128 | R"==(} )==" "\n" |
| 129 | R"==(float gelu_tanh_fwd(float s) { )==" "\n" |
| 130 | R"==(const float sqrt_2_over_pi = 0.79788458347320556640625f; )==" "\n" |
| 131 | R"==(const float fitting_const = 0.044715f; )==" "\n" |
| 132 | R"==(const float g = sqrt_2_over_pi * s * (1.f + fitting_const * s * s); )==" "\n" |
| 133 | R"==(return (0.5f * s * (1.f + tanh_fwd(g))); )==" "\n" |
| 134 | R"==(} )==" "\n" |
| 135 | R"==(float gelu_tanh_bwd(float dd, float s) { )==" "\n" |
| 136 | R"==(const float sqrt_2_over_pi = 0.79788458347320556640625f; )==" "\n" |
| 137 | R"==(const float fitting_const = 0.044715f; )==" "\n" |
| 138 | R"==(const float g = sqrt_2_over_pi * s * (1.f + fitting_const * s * s); )==" "\n" |
| 139 | R"==(const float dg = sqrt_2_over_pi * (1.f + 3.f * fitting_const * s * s); )==" "\n" |
| 140 | R"==(const float v = tanh_fwd(g); )==" "\n" |
| 141 | R"==(return dd * 0.5f * (1.f + v) * (1.f + s * (1.f - v) * dg); )==" "\n" |
| 142 | R"==(} )==" "\n" |
| 143 | R"==(float swish_fwd(float s, float alpha) { )==" "\n" |
| 144 | R"==(float w = -alpha * s; )==" "\n" |
| 145 | R"==(return s / (1.0f + exp(w)); )==" "\n" |
| 146 | R"==(} )==" "\n" |
| 147 | R"==(float swish_bwd(float dd, float s, float alpha) { )==" "\n" |
| 148 | R"==(float v = logistic_fwd(alpha * s); )==" "\n" |
| 149 | R"==(return dd * (v + s * alpha * v * (1.0f - v)); )==" "\n" |
| 150 | R"==(} )==" "\n" |
| 151 | R"==(float log_fwd(float s) { )==" "\n" |
| 152 | R"==(return log(s); )==" "\n" |
| 153 | R"==(} )==" "\n" |
| 154 | R"==(float log_bwd(float dd, float s) { )==" "\n" |
| 155 | R"==(return dd / s; )==" "\n" |
| 156 | R"==(} )==" "\n" |
| 157 | R"==(float clip_fwd(float s, float alpha, float beta) { )==" "\n" |
| 158 | R"==(s = s > alpha ? s : alpha; )==" "\n" |
| 159 | R"==(return s > beta ? beta : s; )==" "\n" |
| 160 | R"==(} )==" "\n" |
| 161 | R"==(float clip_bwd(float dd, float s, float alpha, float beta) { )==" "\n" |
| 162 | R"==(return dd * (alpha < s && s <= beta ? 1 : 0); )==" "\n" |
| 163 | R"==(} )==" "\n" |
| 164 | R"==(float clip_v2_fwd(float s, float alpha, float beta) { )==" "\n" |
| 165 | R"==(s = s > alpha ? s : alpha; )==" "\n" |
| 166 | R"==(return s < beta ? s : beta; )==" "\n" |
| 167 | R"==(} )==" "\n" |
| 168 | R"==(float clip_v2_bwd(float dd, float s, float alpha, float beta) { )==" "\n" |
| 169 | R"==(return dd * (alpha < s && s < beta ? 1 : 0); )==" "\n" |
| 170 | R"==(} )==" "\n" |
| 171 | R"==(float clip_v2_bwd_use_dst(float dd, float d, float alpha, float beta) { )==" "\n" |
| 172 | R"==(return dd * (alpha < d && d < beta ? 1 : 0); )==" "\n" |
| 173 | R"==(} )==" "\n" |
| 174 | R"==(float pow_fwd(float s, float alpha, float beta) { )==" "\n" |
| 175 | R"==(return alpha * pow(s, beta); )==" "\n" |
| 176 | R"==(} )==" "\n" |
| 177 | R"==(float pow_bwd(float dd, float s, float alpha, float beta) { )==" "\n" |
| 178 | R"==(if (beta == 0) return 0; )==" "\n" |
| 179 | R"==(float v = pow_fwd(s, alpha * beta, beta - 1); )==" "\n" |
| 180 | R"==(return dd * v; )==" "\n" |
| 181 | R"==(} )==" "\n" |
| 182 | R"==(float gelu_erf_fwd(float s) { )==" "\n" |
| 183 | R"==(const float sqrt_2_over_2 = 0.707106769084930419921875f; )==" "\n" |
| 184 | R"==(float v = s * sqrt_2_over_2; )==" "\n" |
| 185 | R"==(return 0.5f * s * (1.f + erf(v)); )==" "\n" |
| 186 | R"==(} )==" "\n" |
| 187 | R"==(float gelu_erf_bwd(float dd, float s) { )==" "\n" |
| 188 | R"==(const float two_over_sqrt_pi = 1.12837922573089599609375f; )==" "\n" |
| 189 | R"==(const float sqrt_2_over_2 = 0.707106769084930419921875f; )==" "\n" |
| 190 | R"==(float v = s * sqrt_2_over_2; )==" "\n" |
| 191 | R"==(return dd * 0.5f * (1.f + erf(v) + v * two_over_sqrt_pi * exp(-v * v)); )==" "\n" |
| 192 | R"==(} )==" "\n" |
| 193 | R"==(float round_fwd(float s) { )==" "\n" |
| 194 | R"==(return (float)rint((float)s); )==" "\n" |
| 195 | R"==(} )==" "\n" |
| 196 | R"==(float hardsigmoid_fwd(float s, float alpha, float beta) { )==" "\n" |
| 197 | R"==(float v = alpha * s + beta; )==" "\n" |
| 198 | R"==(return v <= 0.f ? 0.f : v >= 1.f ? 1.f : v; )==" "\n" |
| 199 | R"==(} )==" "\n" |
| 200 | R"==(float hardsigmoid_bwd(float dd, float s, float alpha, float beta) { )==" "\n" |
| 201 | R"==(float v = alpha * s + beta; )==" "\n" |
| 202 | R"==(return v <= 0.f ? 0.f : v >= 1.f ? 0.f : dd * alpha; )==" "\n" |
| 203 | R"==(} )==" "\n" |
| 204 | R"==(float hardswish_fwd(float s, float alpha, float beta) { )==" "\n" |
| 205 | R"==(return s * hardsigmoid_fwd(s, alpha, beta); )==" "\n" |
| 206 | R"==(} )==" "\n" |
| 207 | R"==(float hardswish_bwd(float dd, float s, float alpha, float beta) { )==" "\n" |
| 208 | R"==(float v = alpha * s + beta; )==" "\n" |
| 209 | R"==(float w = 2.f * alpha * s + beta; )==" "\n" |
| 210 | R"==(return (v <= 0.f ? 0.f : v >= 1.f ? dd : dd * w); )==" "\n" |
| 211 | R"==(} )==" "\n" |
| 212 | R"==(float fwd_eltwise_common( )==" "\n" |
| 213 | R"==(int eltwise_alg, float x, float alpha_, float beta_, float scale_) { )==" "\n" |
| 214 | R"==(switch (eltwise_alg) { )==" "\n" |
| 215 | R"==(case RELU: return scale_ * relu_fwd(x, alpha_); break; )==" "\n" |
| 216 | R"==(case LINEAR: return scale_ * linear_fwd(x, alpha_, beta_); break; )==" "\n" |
| 217 | R"==(case SOFT_RELU: return scale_ * soft_relu_fwd(x, alpha_); break; )==" "\n" |
| 218 | R"==(case MISH: return scale_ * mish_fwd(x); break; )==" "\n" |
| 219 | R"==(case LOGISTIC: return scale_ * logistic_fwd(x); break; )==" "\n" |
| 220 | R"==(case TANH: return scale_ * tanh_fwd(x); break; )==" "\n" |
| 221 | R"==(case ELU: return scale_ * elu_fwd(x, alpha_); break; )==" "\n" |
| 222 | R"==(case SQUARE: return scale_ * square_fwd(x); break; )==" "\n" |
| 223 | R"==(case SQRT: return scale_ * sqrt_fwd(x); break; )==" "\n" |
| 224 | R"==(case ABS: return scale_ * abs_fwd(x); break; )==" "\n" |
| 225 | R"==(case EXP: return scale_ * exp_fwd(x); break; )==" "\n" |
| 226 | R"==(case GELU_TANH: return scale_ * gelu_tanh_fwd(x); break; )==" "\n" |
| 227 | R"==(case SWISH: return scale_ * swish_fwd(x, alpha_); break; )==" "\n" |
| 228 | R"==(case LOG: return scale_ * log_fwd(x); break; )==" "\n" |
| 229 | R"==(case CLIP: return scale_ * clip_fwd(x, alpha_, beta_); break; )==" "\n" |
| 230 | R"==(case CLIP_V2: return scale_ * clip_v2_fwd(x, alpha_, beta_); break; )==" "\n" |
| 231 | R"==(case POW: return scale_ * pow_fwd(x, alpha_, beta_); break; )==" "\n" |
| 232 | R"==(case GELU_ERF: return scale_ * gelu_erf_fwd(x); break; )==" "\n" |
| 233 | R"==(case ROUND: return scale_ * round_fwd(x); break; )==" "\n" |
| 234 | R"==(case HARDSWISH: return scale_ * hardswish_fwd(x, alpha_, beta_); break; )==" "\n" |
| 235 | R"==(case HARDSIGMOID: )==" "\n" |
| 236 | R"==(return scale_ * hardsigmoid_fwd(x, alpha_, beta_); )==" "\n" |
| 237 | R"==(break; )==" "\n" |
| 238 | R"==(case RELU_DST: return scale_ * relu_fwd(x, alpha_); break; )==" "\n" |
| 239 | R"==(case LOGISTIC_DST: return scale_ * logistic_fwd(x); break; )==" "\n" |
| 240 | R"==(case TANH_DST: return scale_ * tanh_fwd(x); break; )==" "\n" |
| 241 | R"==(case ELU_DST: return scale_ * elu_fwd(x, alpha_); break; )==" "\n" |
| 242 | R"==(case SQRT_DST: return scale_ * sqrt_fwd(x); break; )==" "\n" |
| 243 | R"==(case EXP_DST: return scale_ * exp_fwd(x); break; )==" "\n" |
| 244 | R"==(case CLIP_V2_DST: return scale_ * clip_v2_fwd(x, alpha_, beta_); break; )==" "\n" |
| 245 | R"==(default: return x; break; )==" "\n" |
| 246 | R"==(} )==" "\n" |
| 247 | R"==(} )==" "\n" |
| 248 | R"==(float fwd_eltwise(float x, float alpha_, float beta_, float scale_) { )==" "\n" |
| 249 | R"==(#ifdef ELTWISE_ALG )==" "\n" |
| 250 | R"==(return fwd_eltwise_common(ELTWISE_ALG, x, alpha_, beta_, scale_); )==" "\n" |
| 251 | R"==(#else )==" "\n" |
| 252 | R"==(return x; )==" "\n" |
| 253 | R"==(#endif )==" "\n" |
| 254 | R"==(} )==" "\n" |
| 255 | R"==(float bwd_eltwise(float x, float y, float alpha_, float beta_) { )==" "\n" |
| 256 | R"==(#ifdef ELTWISE_ALG )==" "\n" |
| 257 | R"==(switch (ELTWISE_ALG) { )==" "\n" |
| 258 | R"==(case RELU: return relu_bwd(x, y, alpha_); break; )==" "\n" |
| 259 | R"==(case LINEAR: return linear_bwd(x, alpha_); break; )==" "\n" |
| 260 | R"==(case SOFT_RELU: return soft_relu_bwd(x, y, alpha_); break; )==" "\n" |
| 261 | R"==(case MISH: return mish_bwd(x, y); break; )==" "\n" |
| 262 | R"==(case LOGISTIC: return logistic_bwd(x, y); break; )==" "\n" |
| 263 | R"==(case TANH: return tanh_bwd(x, y); break; )==" "\n" |
| 264 | R"==(case ELU: return elu_bwd(x, y, alpha_); break; )==" "\n" |
| 265 | R"==(case SQUARE: return square_bwd(x, y); break; )==" "\n" |
| 266 | R"==(case SQRT: return sqrt_bwd(x, y); break; )==" "\n" |
| 267 | R"==(case ABS: return abs_bwd(x, y); break; )==" "\n" |
| 268 | R"==(case EXP: return exp_bwd(x, y); break; )==" "\n" |
| 269 | R"==(case GELU_TANH: return gelu_tanh_bwd(x, y); break; )==" "\n" |
| 270 | R"==(case SWISH: return swish_bwd(x, y, alpha_); break; )==" "\n" |
| 271 | R"==(case LOG: return log_bwd(x, y); break; )==" "\n" |
| 272 | R"==(case CLIP: return clip_bwd(x, y, alpha_, beta_); break; )==" "\n" |
| 273 | R"==(case CLIP_V2: return clip_v2_bwd(x, y, alpha_, beta_); break; )==" "\n" |
| 274 | R"==(case POW: return pow_bwd(x, y, alpha_, beta_); break; )==" "\n" |
| 275 | R"==(case GELU_ERF: return gelu_erf_bwd(x, y); break; )==" "\n" |
| 276 | R"==(case HARDSWISH: return hardswish_bwd(x, y, alpha_, beta_); break; )==" "\n" |
| 277 | R"==(case HARDSIGMOID: return hardsigmoid_bwd(x, y, alpha_, beta_); break; )==" "\n" |
| 278 | R"==(case RELU_DST: return relu_bwd_use_dst(x, y, alpha_); break; )==" "\n" |
| 279 | R"==(case LOGISTIC_DST: return logistic_bwd_use_dst(x, y); break; )==" "\n" |
| 280 | R"==(case TANH_DST: return tanh_bwd_use_dst(x, y); break; )==" "\n" |
| 281 | R"==(case ELU_DST: return elu_bwd_use_dst(x, y, alpha_); break; )==" "\n" |
| 282 | R"==(case SQRT_DST: return sqrt_bwd_use_dst(x, y); break; )==" "\n" |
| 283 | R"==(case EXP_DST: return exp_bwd_use_dst(x, y); break; )==" "\n" |
| 284 | R"==(case CLIP_V2_DST: )==" "\n" |
| 285 | R"==(return clip_v2_bwd_use_dst(x, y, alpha_, beta_); )==" "\n" |
| 286 | R"==(break; )==" "\n" |
| 287 | R"==(default: return x; break; )==" "\n" |
| 288 | R"==(} )==" "\n" |
| 289 | R"==(#else )==" "\n" |
| 290 | R"==(return x; )==" "\n" |
| 291 | R"==(#endif )==" "\n" |
| 292 | R"==(} )==" "\n" |
| 293 | R"==(#endif )==" "\n" |
| 294 | R"==(#endif )==" "\n" |
| 295 | R"==()=="; |
| 296 | } |
| 297 | } |
| 298 | } |
| 299 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.