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_x3(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_x3( 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 >= 3 * sizeof(float); n -= 3 * sizeof(float)) { 43 float vx0 = x[0]; 44 float vx1 = x[1]; 45 float vx2 = x[2]; 46 x += 3; 47 48 const float vz0 = vx0 * vprescale; 49 const float vz1 = vx1 * vprescale; 50 const float vz2 = vx2 * vprescale; 51 52 float vn0 = vz0 * vlog2e + vmagic_bias; 53 float vn1 = vz1 * vlog2e + vmagic_bias; 54 float vn2 = vz2 * vlog2e + vmagic_bias; 55 56 float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23); 57 vn0 -= vmagic_bias; 58 float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23); 59 vn1 -= vmagic_bias; 60 float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23); 61 vn2 -= vmagic_bias; 62 63 float vt0 = vn0 * vminus_ln2_hi + vz0; 64 float vt1 = vn1 * vminus_ln2_hi + vz1; 65 float vt2 = vn2 * vminus_ln2_hi + vz2; 66 67 vt0 = vn0 * vminus_ln2_lo + vt0; 68 vt1 = vn1 * vminus_ln2_lo + vt1; 69 vt2 = vn2 * vminus_ln2_lo + vt2; 70 71 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 72 vs0 = 0.0f; 73 vt0 = 0.0f; 74 } 75 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 76 vs1 = 0.0f; 77 vt1 = 0.0f; 78 } 79 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 80 vs2 = 0.0f; 81 vt2 = 0.0f; 82 } 83 84 float vp0 = vc6 * vt0 + vc5; 85 float vp1 = vc6 * vt1 + vc5; 86 float vp2 = vc6 * vt2 + vc5; 87 88 vp0 = vp0 * vt0 + vc4; 89 vp1 = vp1 * vt1 + vc4; 90 vp2 = vp2 * vt2 + vc4; 91 92 vp0 = vp0 * vt0 + vc3; 93 vp1 = vp1 * vt1 + vc3; 94 vp2 = vp2 * vt2 + vc3; 95 96 vp0 = vp0 * vt0 + vc2; 97 vp1 = vp1 * vt1 + vc2; 98 vp2 = vp2 * vt2 + vc2; 99 100 vp0 *= vt0; 101 vp1 *= vt1; 102 vp2 *= vt2; 103 104 vt0 *= vs0; 105 vs0 -= vone; 106 vt1 *= vs1; 107 vs1 -= vone; 108 vt2 *= vs2; 109 vs2 -= vone; 110 111 vp0 = vp0 * vt0 + vt0; 112 vp1 = vp1 * vt1 + vt1; 113 vp2 = vp2 * vt2 + vt2; 114 115 const float ve0 = (vp0 + vs0) * valpha; 116 float vy0 = vx0 * vbeta; 117 const float ve1 = (vp1 + vs1) * valpha; 118 float vy1 = vx1 * vbeta; 119 const float ve2 = (vp2 + vs2) * valpha; 120 float vy2 = vx2 * vbeta; 121 122 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 123 vy0 = ve0; 124 } 125 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 126 vy1 = ve1; 127 } 128 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 129 vy2 = ve2; 130 } 131 132 y[0] = vy0; 133 y[1] = vy1; 134 y[2] = vy2; 135 y += 3; 136 } 137 if XNN_UNLIKELY(n != 0) { 138 do { 139 float vx = *x++; 140 141 const float vz = vx * vprescale; 142 143 float vn = vz * vlog2e + vmagic_bias; 144 float vs = fp32_from_bits(fp32_to_bits(vn) << 23); 145 vn -= vmagic_bias; 146 147 float vt = vn * vminus_ln2_hi + vz; 148 vt = vn * vminus_ln2_lo + vt; 149 150 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 151 vs = 0.0f; 152 vt = 0.0f; 153 } 154 155 float vp = vc6 * vt + vc5; 156 vp = vp * vt + vc4; 157 vp = vp * vt + vc3; 158 vp = vp * vt + vc2; 159 vp *= vt; 160 161 vt *= vs; 162 vs -= vone; 163 vp = vp * vt + vt; 164 const float ve = (vp + vs) * valpha; 165 166 float vy = vx * vbeta; 167 if XNN_UNPREDICTABLE(vx < 0.0f) { 168 vy = ve; 169 } 170 171 *y++ = vy; 172 173 n -= sizeof(float); 174 } while (n != 0); 175 } 176 } 177