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_x2(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_x2( 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 >= 2 * sizeof(float); n -= 2 * sizeof(float)) { 43 float vx0 = x[0]; 44 float vx1 = x[1]; 45 x += 2; 46 47 const float vz0 = vx0 * vprescale; 48 const float vz1 = vx1 * vprescale; 49 50 float vn0 = vz0 * vlog2e + vmagic_bias; 51 float vn1 = vz1 * vlog2e + vmagic_bias; 52 53 float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23); 54 vn0 -= vmagic_bias; 55 float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23); 56 vn1 -= vmagic_bias; 57 58 float vt0 = vn0 * vminus_ln2_hi + vz0; 59 float vt1 = vn1 * vminus_ln2_hi + vz1; 60 61 vt0 = vn0 * vminus_ln2_lo + vt0; 62 vt1 = vn1 * vminus_ln2_lo + vt1; 63 64 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 65 vs0 = 0.0f; 66 vt0 = 0.0f; 67 } 68 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 69 vs1 = 0.0f; 70 vt1 = 0.0f; 71 } 72 73 float vp0 = vc6 * vt0 + vc5; 74 float vp1 = vc6 * vt1 + vc5; 75 76 vp0 = vp0 * vt0 + vc4; 77 vp1 = vp1 * vt1 + vc4; 78 79 vp0 = vp0 * vt0 + vc3; 80 vp1 = vp1 * vt1 + vc3; 81 82 vp0 = vp0 * vt0 + vc2; 83 vp1 = vp1 * vt1 + vc2; 84 85 vp0 *= vt0; 86 vp1 *= vt1; 87 88 vt0 *= vs0; 89 vs0 -= vone; 90 vt1 *= vs1; 91 vs1 -= vone; 92 93 vp0 = vp0 * vt0 + vt0; 94 vp1 = vp1 * vt1 + vt1; 95 96 const float ve0 = (vp0 + vs0) * valpha; 97 float vy0 = vx0 * vbeta; 98 const float ve1 = (vp1 + vs1) * valpha; 99 float vy1 = vx1 * vbeta; 100 101 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 102 vy0 = ve0; 103 } 104 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 105 vy1 = ve1; 106 } 107 108 y[0] = vy0; 109 y[1] = vy1; 110 y += 2; 111 } 112 if XNN_UNLIKELY(n != 0) { 113 float vx = *x; 114 115 const float vz = vx * vprescale; 116 117 float vn = vz * vlog2e + vmagic_bias; 118 float vs = fp32_from_bits(fp32_to_bits(vn) << 23); 119 vn -= vmagic_bias; 120 121 float vt = vn * vminus_ln2_hi + vz; 122 vt = vn * vminus_ln2_lo + vt; 123 124 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 125 vs = 0.0f; 126 vt = 0.0f; 127 } 128 129 float vp = vc6 * vt + vc5; 130 vp = vp * vt + vc4; 131 vp = vp * vt + vc3; 132 vp = vp * vt + vc2; 133 vp *= vt; 134 135 vt *= vs; 136 vs -= vone; 137 vp = vp * vt + vt; 138 const float ve = (vp + vs) * valpha; 139 140 float vy = vx * vbeta; 141 if XNN_UNPREDICTABLE(vx < 0.0f) { 142 vy = ve; 143 } 144 145 *y = vy; 146 } 147 } 148