1 // Auto-generated file. Do not edit! 2 // Template: src/f32-velu/scalar-rr2-p6.c.in 3 // Generator: tools/xngen 4 // 5 // Copyright 2020 Google LLC 6 // 7 // This source code is licensed under the BSD-style license found in the 8 // LICENSE file in the root directory of this source tree. 9 10 #include <assert.h> 11 #include <math.h> 12 13 #include <xnnpack/common.h> 14 #include <xnnpack/math.h> 15 #include <xnnpack/vunary.h> 16 17 xnn_f32_velu_ukernel__scalar_rr2_p6_x6(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18void xnn_f32_velu_ukernel__scalar_rr2_p6_x6( 19 size_t n, 20 const float* x, 21 float* y, 22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 23 { 24 assert(n % sizeof(float) == 0); 25 26 const float vprescale = params->scalar_rr2_p6.prescale; 27 const float valpha = params->scalar_rr2_p6.alpha; 28 const float vbeta = params->scalar_rr2_p6.beta; 29 const float vmagic_bias = params->scalar_rr2_p6.magic_bias; 30 const float vlog2e = params->scalar_rr2_p6.log2e; 31 const float vsat_cutoff = params->scalar_rr2_p6.sat_cutoff; 32 const float vminus_ln2_hi = params->scalar_rr2_p6.minus_ln2_hi; 33 const float vminus_ln2_lo = params->scalar_rr2_p6.minus_ln2_lo; 34 const float vc6 = params->scalar_rr2_p6.c6; 35 const float vc5 = params->scalar_rr2_p6.c5; 36 const float vc4 = params->scalar_rr2_p6.c4; 37 const float vc3 = params->scalar_rr2_p6.c3; 38 const float vc2 = params->scalar_rr2_p6.c2; 39 const float vone = params->scalar_rr2_p6.one; 40 41 for (; n >= 6 * sizeof(float); n -= 6 * sizeof(float)) { 42 float vx0 = x[0]; 43 float vx1 = x[1]; 44 float vx2 = x[2]; 45 float vx3 = x[3]; 46 float vx4 = x[4]; 47 float vx5 = x[5]; 48 x += 6; 49 50 const float vz0 = vx0 * vprescale; 51 const float vz1 = vx1 * vprescale; 52 const float vz2 = vx2 * vprescale; 53 const float vz3 = vx3 * vprescale; 54 const float vz4 = vx4 * vprescale; 55 const float vz5 = vx5 * vprescale; 56 57 float vn0 = vz0 * vlog2e + vmagic_bias; 58 float vn1 = vz1 * vlog2e + vmagic_bias; 59 float vn2 = vz2 * vlog2e + vmagic_bias; 60 float vn3 = vz3 * vlog2e + vmagic_bias; 61 float vn4 = vz4 * vlog2e + vmagic_bias; 62 float vn5 = vz5 * vlog2e + vmagic_bias; 63 64 float vs0 = uint32_as_float(float_as_uint32(vn0) << 23); 65 vn0 -= vmagic_bias; 66 float vs1 = uint32_as_float(float_as_uint32(vn1) << 23); 67 vn1 -= vmagic_bias; 68 float vs2 = uint32_as_float(float_as_uint32(vn2) << 23); 69 vn2 -= vmagic_bias; 70 float vs3 = uint32_as_float(float_as_uint32(vn3) << 23); 71 vn3 -= vmagic_bias; 72 float vs4 = uint32_as_float(float_as_uint32(vn4) << 23); 73 vn4 -= vmagic_bias; 74 float vs5 = uint32_as_float(float_as_uint32(vn5) << 23); 75 vn5 -= vmagic_bias; 76 77 float vt0 = vn0 * vminus_ln2_hi + vz0; 78 float vt1 = vn1 * vminus_ln2_hi + vz1; 79 float vt2 = vn2 * vminus_ln2_hi + vz2; 80 float vt3 = vn3 * vminus_ln2_hi + vz3; 81 float vt4 = vn4 * vminus_ln2_hi + vz4; 82 float vt5 = vn5 * vminus_ln2_hi + vz5; 83 84 vt0 = vn0 * vminus_ln2_lo + vt0; 85 vt1 = vn1 * vminus_ln2_lo + vt1; 86 vt2 = vn2 * vminus_ln2_lo + vt2; 87 vt3 = vn3 * vminus_ln2_lo + vt3; 88 vt4 = vn4 * vminus_ln2_lo + vt4; 89 vt5 = vn5 * vminus_ln2_lo + vt5; 90 91 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 92 vs0 = 0.0f; 93 vt0 = 0.0f; 94 } 95 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 96 vs1 = 0.0f; 97 vt1 = 0.0f; 98 } 99 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 100 vs2 = 0.0f; 101 vt2 = 0.0f; 102 } 103 if XNN_UNPREDICTABLE(vz3 <= vsat_cutoff) { 104 vs3 = 0.0f; 105 vt3 = 0.0f; 106 } 107 if XNN_UNPREDICTABLE(vz4 <= vsat_cutoff) { 108 vs4 = 0.0f; 109 vt4 = 0.0f; 110 } 111 if XNN_UNPREDICTABLE(vz5 <= vsat_cutoff) { 112 vs5 = 0.0f; 113 vt5 = 0.0f; 114 } 115 116 float vp0 = vc6 * vt0 + vc5; 117 float vp1 = vc6 * vt1 + vc5; 118 float vp2 = vc6 * vt2 + vc5; 119 float vp3 = vc6 * vt3 + vc5; 120 float vp4 = vc6 * vt4 + vc5; 121 float vp5 = vc6 * vt5 + vc5; 122 123 vp0 = vp0 * vt0 + vc4; 124 vp1 = vp1 * vt1 + vc4; 125 vp2 = vp2 * vt2 + vc4; 126 vp3 = vp3 * vt3 + vc4; 127 vp4 = vp4 * vt4 + vc4; 128 vp5 = vp5 * vt5 + vc4; 129 130 vp0 = vp0 * vt0 + vc3; 131 vp1 = vp1 * vt1 + vc3; 132 vp2 = vp2 * vt2 + vc3; 133 vp3 = vp3 * vt3 + vc3; 134 vp4 = vp4 * vt4 + vc3; 135 vp5 = vp5 * vt5 + vc3; 136 137 vp0 = vp0 * vt0 + vc2; 138 vp1 = vp1 * vt1 + vc2; 139 vp2 = vp2 * vt2 + vc2; 140 vp3 = vp3 * vt3 + vc2; 141 vp4 = vp4 * vt4 + vc2; 142 vp5 = vp5 * vt5 + vc2; 143 144 vp0 *= vt0; 145 vp1 *= vt1; 146 vp2 *= vt2; 147 vp3 *= vt3; 148 vp4 *= vt4; 149 vp5 *= vt5; 150 151 vt0 *= vs0; 152 vs0 -= vone; 153 vt1 *= vs1; 154 vs1 -= vone; 155 vt2 *= vs2; 156 vs2 -= vone; 157 vt3 *= vs3; 158 vs3 -= vone; 159 vt4 *= vs4; 160 vs4 -= vone; 161 vt5 *= vs5; 162 vs5 -= vone; 163 164 vp0 = vp0 * vt0 + vt0; 165 vp1 = vp1 * vt1 + vt1; 166 vp2 = vp2 * vt2 + vt2; 167 vp3 = vp3 * vt3 + vt3; 168 vp4 = vp4 * vt4 + vt4; 169 vp5 = vp5 * vt5 + vt5; 170 171 const float ve0 = (vp0 + vs0) * valpha; 172 float vy0 = vx0 * vbeta; 173 const float ve1 = (vp1 + vs1) * valpha; 174 float vy1 = vx1 * vbeta; 175 const float ve2 = (vp2 + vs2) * valpha; 176 float vy2 = vx2 * vbeta; 177 const float ve3 = (vp3 + vs3) * valpha; 178 float vy3 = vx3 * vbeta; 179 const float ve4 = (vp4 + vs4) * valpha; 180 float vy4 = vx4 * vbeta; 181 const float ve5 = (vp5 + vs5) * valpha; 182 float vy5 = vx5 * vbeta; 183 184 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 185 vy0 = ve0; 186 } 187 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 188 vy1 = ve1; 189 } 190 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 191 vy2 = ve2; 192 } 193 if XNN_UNPREDICTABLE(vx3 < 0.0f) { 194 vy3 = ve3; 195 } 196 if XNN_UNPREDICTABLE(vx4 < 0.0f) { 197 vy4 = ve4; 198 } 199 if XNN_UNPREDICTABLE(vx5 < 0.0f) { 200 vy5 = ve5; 201 } 202 203 y[0] = vy0; 204 y[1] = vy1; 205 y[2] = vy2; 206 y[3] = vy3; 207 y[4] = vy4; 208 y[5] = vy5; 209 y += 6; 210 } 211 if XNN_UNLIKELY(n != 0) { 212 do { 213 float vx = *x++; 214 215 const float vz = vx * vprescale; 216 217 float vn = vz * vlog2e + vmagic_bias; 218 float vs = uint32_as_float(float_as_uint32(vn) << 23); 219 vn -= vmagic_bias; 220 221 float vt = vn * vminus_ln2_hi + vz; 222 vt = vn * vminus_ln2_lo + vt; 223 224 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 225 vs = 0.0f; 226 vt = 0.0f; 227 } 228 229 float vp = vc6 * vt + vc5; 230 vp = vp * vt + vc4; 231 vp = vp * vt + vc3; 232 vp = vp * vt + vc2; 233 vp *= vt; 234 235 vt *= vs; 236 vs -= vone; 237 vp = vp * vt + vt; 238 const float ve = (vp + vs) * valpha; 239 240 float vy = vx * vbeta; 241 if XNN_UNPREDICTABLE(vx < 0.0f) { 242 vy = ve; 243 } 244 245 *y++ = vy; 246 247 n -= sizeof(float); 248 } while (n != 0); 249 } 250 } 251