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.prescale; 28 const float valpha = params->scalar.alpha; 29 const float vbeta = params->scalar.beta; 30 31 const float vmagic_bias = 0x1.8000FEp23f; 32 const float vlog2e = 0x1.715476p+0f; 33 const float vsat_cutoff = -0x1.154246p+4f; 34 const float vminus_ln2_hi = -0x1.62E440p-1f; 35 const float vminus_ln2_lo = 0x1.0105C6p-21f; 36 const float vc6 = 0x1.6b7338p-10f; 37 const float vc5 = 0x1.12278Ep-7f; 38 const float vc4 = 0x1.555716p-5f; 39 const float vc3 = 0x1.5554B0p-3f; 40 const float vc2 = 0x1.FFFFFEp-2f; 41 const float vone = 1.0f; 42 43 for (; n >= 3 * sizeof(float); n -= 3 * sizeof(float)) { 44 float vx0 = x[0]; 45 float vx1 = x[1]; 46 float vx2 = x[2]; 47 x += 3; 48 49 const float vz0 = vx0 * vprescale; 50 const float vz1 = vx1 * vprescale; 51 const float vz2 = vx2 * vprescale; 52 53 float vn0 = vz0 * vlog2e + vmagic_bias; 54 float vn1 = vz1 * vlog2e + vmagic_bias; 55 float vn2 = vz2 * vlog2e + vmagic_bias; 56 57 float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23); 58 vn0 -= vmagic_bias; 59 float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23); 60 vn1 -= vmagic_bias; 61 float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23); 62 vn2 -= vmagic_bias; 63 64 float vt0 = vn0 * vminus_ln2_hi + vz0; 65 float vt1 = vn1 * vminus_ln2_hi + vz1; 66 float vt2 = vn2 * vminus_ln2_hi + vz2; 67 68 vt0 = vn0 * vminus_ln2_lo + vt0; 69 vt1 = vn1 * vminus_ln2_lo + vt1; 70 vt2 = vn2 * vminus_ln2_lo + vt2; 71 72 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 73 vs0 = 0.0f; 74 vt0 = 0.0f; 75 } 76 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 77 vs1 = 0.0f; 78 vt1 = 0.0f; 79 } 80 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 81 vs2 = 0.0f; 82 vt2 = 0.0f; 83 } 84 85 float vp0 = vc6 * vt0 + vc5; 86 float vp1 = vc6 * vt1 + vc5; 87 float vp2 = vc6 * vt2 + vc5; 88 89 vp0 = vp0 * vt0 + vc4; 90 vp1 = vp1 * vt1 + vc4; 91 vp2 = vp2 * vt2 + vc4; 92 93 vp0 = vp0 * vt0 + vc3; 94 vp1 = vp1 * vt1 + vc3; 95 vp2 = vp2 * vt2 + vc3; 96 97 vp0 = vp0 * vt0 + vc2; 98 vp1 = vp1 * vt1 + vc2; 99 vp2 = vp2 * vt2 + vc2; 100 101 vp0 *= vt0; 102 vp1 *= vt1; 103 vp2 *= vt2; 104 105 vt0 *= vs0; 106 vs0 -= vone; 107 vt1 *= vs1; 108 vs1 -= vone; 109 vt2 *= vs2; 110 vs2 -= vone; 111 112 vp0 = vp0 * vt0 + vt0; 113 vp1 = vp1 * vt1 + vt1; 114 vp2 = vp2 * vt2 + vt2; 115 116 const float ve0 = (vp0 + vs0) * valpha; 117 float vy0 = vx0 * vbeta; 118 const float ve1 = (vp1 + vs1) * valpha; 119 float vy1 = vx1 * vbeta; 120 const float ve2 = (vp2 + vs2) * valpha; 121 float vy2 = vx2 * vbeta; 122 123 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 124 vy0 = ve0; 125 } 126 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 127 vy1 = ve1; 128 } 129 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 130 vy2 = ve2; 131 } 132 133 y[0] = vy0; 134 y[1] = vy1; 135 y[2] = vy2; 136 y += 3; 137 } 138 if XNN_UNLIKELY(n != 0) { 139 do { 140 float vx = *x++; 141 142 const float vz = vx * vprescale; 143 144 float vn = vz * vlog2e + vmagic_bias; 145 float vs = fp32_from_bits(fp32_to_bits(vn) << 23); 146 vn -= vmagic_bias; 147 148 float vt = vn * vminus_ln2_hi + vz; 149 vt = vn * vminus_ln2_lo + vt; 150 151 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 152 vs = 0.0f; 153 vt = 0.0f; 154 } 155 156 float vp = vc6 * vt + vc5; 157 vp = vp * vt + vc4; 158 vp = vp * vt + vc3; 159 vp = vp * vt + vc2; 160 vp *= vt; 161 162 vt *= vs; 163 vs -= vone; 164 vp = vp * vt + vt; 165 const float ve = (vp + vs) * valpha; 166 167 float vy = vx * vbeta; 168 if XNN_UNPREDICTABLE(vx < 0.0f) { 169 vy = ve; 170 } 171 172 *y++ = vy; 173 174 n -= sizeof(float); 175 } while (n != 0); 176 } 177 } 178