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_x3(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_x3( 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 >= 3 * sizeof(float); n -= 3 * sizeof(float)) { 42 float vx0 = x[0]; 43 float vx1 = x[1]; 44 float vx2 = x[2]; 45 x += 3; 46 47 const float vz0 = vx0 * vprescale; 48 const float vz1 = vx1 * vprescale; 49 const float vz2 = vx2 * vprescale; 50 51 float vn0 = vz0 * vlog2e + vmagic_bias; 52 float vn1 = vz1 * vlog2e + vmagic_bias; 53 float vn2 = vz2 * vlog2e + vmagic_bias; 54 55 float vs0 = uint32_as_float(float_as_uint32(vn0) << 23); 56 vn0 -= vmagic_bias; 57 float vs1 = uint32_as_float(float_as_uint32(vn1) << 23); 58 vn1 -= vmagic_bias; 59 float vs2 = uint32_as_float(float_as_uint32(vn2) << 23); 60 vn2 -= vmagic_bias; 61 62 float vt0 = vn0 * vminus_ln2_hi + vz0; 63 float vt1 = vn1 * vminus_ln2_hi + vz1; 64 float vt2 = vn2 * vminus_ln2_hi + vz2; 65 66 vt0 = vn0 * vminus_ln2_lo + vt0; 67 vt1 = vn1 * vminus_ln2_lo + vt1; 68 vt2 = vn2 * vminus_ln2_lo + vt2; 69 70 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 71 vs0 = 0.0f; 72 vt0 = 0.0f; 73 } 74 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 75 vs1 = 0.0f; 76 vt1 = 0.0f; 77 } 78 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 79 vs2 = 0.0f; 80 vt2 = 0.0f; 81 } 82 83 float vp0 = vc6 * vt0 + vc5; 84 float vp1 = vc6 * vt1 + vc5; 85 float vp2 = vc6 * vt2 + vc5; 86 87 vp0 = vp0 * vt0 + vc4; 88 vp1 = vp1 * vt1 + vc4; 89 vp2 = vp2 * vt2 + vc4; 90 91 vp0 = vp0 * vt0 + vc3; 92 vp1 = vp1 * vt1 + vc3; 93 vp2 = vp2 * vt2 + vc3; 94 95 vp0 = vp0 * vt0 + vc2; 96 vp1 = vp1 * vt1 + vc2; 97 vp2 = vp2 * vt2 + vc2; 98 99 vp0 *= vt0; 100 vp1 *= vt1; 101 vp2 *= vt2; 102 103 vt0 *= vs0; 104 vs0 -= vone; 105 vt1 *= vs1; 106 vs1 -= vone; 107 vt2 *= vs2; 108 vs2 -= vone; 109 110 vp0 = vp0 * vt0 + vt0; 111 vp1 = vp1 * vt1 + vt1; 112 vp2 = vp2 * vt2 + vt2; 113 114 const float ve0 = (vp0 + vs0) * valpha; 115 float vy0 = vx0 * vbeta; 116 const float ve1 = (vp1 + vs1) * valpha; 117 float vy1 = vx1 * vbeta; 118 const float ve2 = (vp2 + vs2) * valpha; 119 float vy2 = vx2 * vbeta; 120 121 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 122 vy0 = ve0; 123 } 124 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 125 vy1 = ve1; 126 } 127 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 128 vy2 = ve2; 129 } 130 131 y[0] = vy0; 132 y[1] = vy1; 133 y[2] = vy2; 134 y += 3; 135 } 136 if XNN_UNLIKELY(n != 0) { 137 do { 138 float vx = *x++; 139 140 const float vz = vx * vprescale; 141 142 float vn = vz * vlog2e + vmagic_bias; 143 float vs = uint32_as_float(float_as_uint32(vn) << 23); 144 vn -= vmagic_bias; 145 146 float vt = vn * vminus_ln2_hi + vz; 147 vt = vn * vminus_ln2_lo + vt; 148 149 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 150 vs = 0.0f; 151 vt = 0.0f; 152 } 153 154 float vp = vc6 * vt + vc5; 155 vp = vp * vt + vc4; 156 vp = vp * vt + vc3; 157 vp = vp * vt + vc2; 158 vp *= vt; 159 160 vt *= vs; 161 vs -= vone; 162 vp = vp * vt + vt; 163 const float ve = (vp + vs) * valpha; 164 165 float vy = vx * vbeta; 166 if XNN_UNPREDICTABLE(vx < 0.0f) { 167 vy = ve; 168 } 169 170 *y++ = vy; 171 172 n -= sizeof(float); 173 } while (n != 0); 174 } 175 } 176