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/vunary.h> 15 16 #include <fp16/bitcasts.h> 17 18 xnn_f32_velu_ukernel__scalar_rr2_p6_x4(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19void xnn_f32_velu_ukernel__scalar_rr2_p6_x4( 20 size_t n, 21 const float* x, 22 float* y, 23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 24 { 25 assert(n % sizeof(float) == 0); 26 27 const float vprescale = params->scalar_rr2_p6.prescale; 28 const float valpha = params->scalar_rr2_p6.alpha; 29 const float vbeta = params->scalar_rr2_p6.beta; 30 const float vmagic_bias = params->scalar_rr2_p6.magic_bias; 31 const float vlog2e = params->scalar_rr2_p6.log2e; 32 const float vsat_cutoff = params->scalar_rr2_p6.sat_cutoff; 33 const float vminus_ln2_hi = params->scalar_rr2_p6.minus_ln2_hi; 34 const float vminus_ln2_lo = params->scalar_rr2_p6.minus_ln2_lo; 35 const float vc6 = params->scalar_rr2_p6.c6; 36 const float vc5 = params->scalar_rr2_p6.c5; 37 const float vc4 = params->scalar_rr2_p6.c4; 38 const float vc3 = params->scalar_rr2_p6.c3; 39 const float vc2 = params->scalar_rr2_p6.c2; 40 const float vone = params->scalar_rr2_p6.one; 41 42 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 43 float vx0 = x[0]; 44 float vx1 = x[1]; 45 float vx2 = x[2]; 46 float vx3 = x[3]; 47 x += 4; 48 49 const float vz0 = vx0 * vprescale; 50 const float vz1 = vx1 * vprescale; 51 const float vz2 = vx2 * vprescale; 52 const float vz3 = vx3 * vprescale; 53 54 float vn0 = vz0 * vlog2e + vmagic_bias; 55 float vn1 = vz1 * vlog2e + vmagic_bias; 56 float vn2 = vz2 * vlog2e + vmagic_bias; 57 float vn3 = vz3 * vlog2e + vmagic_bias; 58 59 float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23); 60 vn0 -= vmagic_bias; 61 float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23); 62 vn1 -= vmagic_bias; 63 float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23); 64 vn2 -= vmagic_bias; 65 float vs3 = fp32_from_bits(fp32_to_bits(vn3) << 23); 66 vn3 -= vmagic_bias; 67 68 float vt0 = vn0 * vminus_ln2_hi + vz0; 69 float vt1 = vn1 * vminus_ln2_hi + vz1; 70 float vt2 = vn2 * vminus_ln2_hi + vz2; 71 float vt3 = vn3 * vminus_ln2_hi + vz3; 72 73 vt0 = vn0 * vminus_ln2_lo + vt0; 74 vt1 = vn1 * vminus_ln2_lo + vt1; 75 vt2 = vn2 * vminus_ln2_lo + vt2; 76 vt3 = vn3 * vminus_ln2_lo + vt3; 77 78 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 79 vs0 = 0.0f; 80 vt0 = 0.0f; 81 } 82 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 83 vs1 = 0.0f; 84 vt1 = 0.0f; 85 } 86 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 87 vs2 = 0.0f; 88 vt2 = 0.0f; 89 } 90 if XNN_UNPREDICTABLE(vz3 <= vsat_cutoff) { 91 vs3 = 0.0f; 92 vt3 = 0.0f; 93 } 94 95 float vp0 = vc6 * vt0 + vc5; 96 float vp1 = vc6 * vt1 + vc5; 97 float vp2 = vc6 * vt2 + vc5; 98 float vp3 = vc6 * vt3 + vc5; 99 100 vp0 = vp0 * vt0 + vc4; 101 vp1 = vp1 * vt1 + vc4; 102 vp2 = vp2 * vt2 + vc4; 103 vp3 = vp3 * vt3 + vc4; 104 105 vp0 = vp0 * vt0 + vc3; 106 vp1 = vp1 * vt1 + vc3; 107 vp2 = vp2 * vt2 + vc3; 108 vp3 = vp3 * vt3 + vc3; 109 110 vp0 = vp0 * vt0 + vc2; 111 vp1 = vp1 * vt1 + vc2; 112 vp2 = vp2 * vt2 + vc2; 113 vp3 = vp3 * vt3 + vc2; 114 115 vp0 *= vt0; 116 vp1 *= vt1; 117 vp2 *= vt2; 118 vp3 *= vt3; 119 120 vt0 *= vs0; 121 vs0 -= vone; 122 vt1 *= vs1; 123 vs1 -= vone; 124 vt2 *= vs2; 125 vs2 -= vone; 126 vt3 *= vs3; 127 vs3 -= vone; 128 129 vp0 = vp0 * vt0 + vt0; 130 vp1 = vp1 * vt1 + vt1; 131 vp2 = vp2 * vt2 + vt2; 132 vp3 = vp3 * vt3 + vt3; 133 134 const float ve0 = (vp0 + vs0) * valpha; 135 float vy0 = vx0 * vbeta; 136 const float ve1 = (vp1 + vs1) * valpha; 137 float vy1 = vx1 * vbeta; 138 const float ve2 = (vp2 + vs2) * valpha; 139 float vy2 = vx2 * vbeta; 140 const float ve3 = (vp3 + vs3) * valpha; 141 float vy3 = vx3 * vbeta; 142 143 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 144 vy0 = ve0; 145 } 146 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 147 vy1 = ve1; 148 } 149 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 150 vy2 = ve2; 151 } 152 if XNN_UNPREDICTABLE(vx3 < 0.0f) { 153 vy3 = ve3; 154 } 155 156 y[0] = vy0; 157 y[1] = vy1; 158 y[2] = vy2; 159 y[3] = vy3; 160 y += 4; 161 } 162 if XNN_UNLIKELY(n != 0) { 163 do { 164 float vx = *x++; 165 166 const float vz = vx * vprescale; 167 168 float vn = vz * vlog2e + vmagic_bias; 169 float vs = fp32_from_bits(fp32_to_bits(vn) << 23); 170 vn -= vmagic_bias; 171 172 float vt = vn * vminus_ln2_hi + vz; 173 vt = vn * vminus_ln2_lo + vt; 174 175 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 176 vs = 0.0f; 177 vt = 0.0f; 178 } 179 180 float vp = vc6 * vt + vc5; 181 vp = vp * vt + vc4; 182 vp = vp * vt + vc3; 183 vp = vp * vt + vc2; 184 vp *= vt; 185 186 vt *= vs; 187 vs -= vone; 188 vp = vp * vt + vt; 189 const float ve = (vp + vs) * valpha; 190 191 float vy = vx * vbeta; 192 if XNN_UNPREDICTABLE(vx < 0.0f) { 193 vy = ve; 194 } 195 196 *y++ = vy; 197 198 n -= sizeof(float); 199 } while (n != 0); 200 } 201 } 202