1 // Auto-generated file. Do not edit! 2 // Template: src/f32-velu/scalar-rr2-lut16-p3.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 19 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_16[16]; 20 xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x3(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])21void xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x3( 22 size_t n, 23 const float* x, 24 float* y, 25 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) 26 { 27 assert(n % sizeof(float) == 0); 28 29 const float vprescale = params->scalar_rr2_lut16_p3.prescale; 30 const float valpha = params->scalar_rr2_lut16_p3.alpha; 31 const float vbeta = params->scalar_rr2_lut16_p3.beta; 32 const float vmagic_bias = params->scalar_rr2_lut16_p3.magic_bias; 33 const float vlog2e = params->scalar_rr2_lut16_p3.log2e; 34 const uint32_t vindex_mask = UINT32_C(0xF); 35 const float vsat_cutoff = params->scalar_rr2_lut16_p3.sat_cutoff; 36 const float vminus_ln2_hi = params->scalar_rr2_lut16_p3.minus_ln2_hi; 37 const float vminus_ln2_lo = params->scalar_rr2_lut16_p3.minus_ln2_lo; 38 const float vc3 = params->scalar_rr2_lut16_p3.c3; 39 const float vc2 = params->scalar_rr2_lut16_p3.c2; 40 const float vone = params->scalar_rr2_lut16_p3.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 const uint32_t ven0 = fp32_to_bits(vn0) << 19; 57 const uint32_t vidx0 = fp32_to_bits(vn0) & vindex_mask; 58 vn0 -= vmagic_bias; 59 const uint32_t ven1 = fp32_to_bits(vn1) << 19; 60 const uint32_t vidx1 = fp32_to_bits(vn1) & vindex_mask; 61 vn1 -= vmagic_bias; 62 const uint32_t ven2 = fp32_to_bits(vn2) << 19; 63 const uint32_t vidx2 = fp32_to_bits(vn2) & vindex_mask; 64 vn2 -= vmagic_bias; 65 66 float vt0 = vn0 * vminus_ln2_hi + vz0; 67 float vs0 = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx0] + ven0); 68 float vt1 = vn1 * vminus_ln2_hi + vz1; 69 float vs1 = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx1] + ven1); 70 float vt2 = vn2 * vminus_ln2_hi + vz2; 71 float vs2 = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx2] + ven2); 72 73 vt0 = vn0 * vminus_ln2_lo + vt0; 74 if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) { 75 vs0 = 0.0f; 76 vt0 = 0.0f; 77 } 78 vt1 = vn1 * vminus_ln2_lo + vt1; 79 if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) { 80 vs1 = 0.0f; 81 vt1 = 0.0f; 82 } 83 vt2 = vn2 * vminus_ln2_lo + vt2; 84 if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) { 85 vs2 = 0.0f; 86 vt2 = 0.0f; 87 } 88 89 float vp0 = vc3 * vt0 + vc2; 90 float vp1 = vc3 * vt1 + vc2; 91 float vp2 = vc3 * vt2 + vc2; 92 93 vp0 *= vt0; 94 vp1 *= vt1; 95 vp2 *= vt2; 96 97 vt0 *= vs0; 98 vs0 -= vone; 99 vt1 *= vs1; 100 vs1 -= vone; 101 vt2 *= vs2; 102 vs2 -= vone; 103 104 vp0 = vp0 * vt0 + vt0; 105 vp1 = vp1 * vt1 + vt1; 106 vp2 = vp2 * vt2 + vt2; 107 108 const float ve0 = (vp0 + vs0) * valpha; 109 float vy0 = vx0 * vbeta; 110 const float ve1 = (vp1 + vs1) * valpha; 111 float vy1 = vx1 * vbeta; 112 const float ve2 = (vp2 + vs2) * valpha; 113 float vy2 = vx2 * vbeta; 114 115 if XNN_UNPREDICTABLE(vx0 < 0.0f) { 116 vy0 = ve0; 117 } 118 if XNN_UNPREDICTABLE(vx1 < 0.0f) { 119 vy1 = ve1; 120 } 121 if XNN_UNPREDICTABLE(vx2 < 0.0f) { 122 vy2 = ve2; 123 } 124 125 y[0] = vy0; 126 y[1] = vy1; 127 y[2] = vy2; 128 y += 3; 129 } 130 if XNN_UNLIKELY(n != 0) { 131 do { 132 float vx = *x++; 133 134 const float vz = vx * vprescale; 135 136 float vn = vz * vlog2e + vmagic_bias; 137 const uint32_t ven = fp32_to_bits(vn) << 19; 138 const uint32_t vidx = fp32_to_bits(vn) & vindex_mask; 139 vn -= vmagic_bias; 140 141 float vt = vn * vminus_ln2_hi + vz; 142 float vs = fp32_from_bits(xnn_table_exp2minus_k_over_16[vidx] + ven); 143 144 vt = vn * vminus_ln2_lo + vt; 145 if XNN_UNPREDICTABLE(vz <= vsat_cutoff) { 146 vs = 0.0f; 147 vt = 0.0f; 148 } 149 150 float vp = vc3 * vt + vc2; 151 vp *= vt; 152 153 vt *= vs; 154 vs -= vone; 155 vp = vp * vt + vt; 156 const float ve = (vp + vs) * valpha; 157 158 float vy = vx * vbeta; 159 if XNN_UNPREDICTABLE(vx < 0.0f) { 160 vy = ve; 161 } 162 163 *y++ = vy; 164 165 n -= sizeof(float); 166 } while (n != 0); 167 } 168 } 169