1 // Auto-generated file. Do not edit! 2 // Template: src/f32-vsigmoid/scalar-rr2-lut64-p2-div.c.in 3 // Generator: tools/xngen 4 // 5 // Copyright 2019 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 18 // Note redefine as uint32[] to avoid redundant bitcasts. 19 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_64[64]; 20 xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])21void xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x4( 22 size_t n, 23 const float* x, 24 float* y, 25 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) 26 { 27 assert(n % sizeof(float) == 0); 28 29 const float vmagic_bias = params->scalar_rr2_lut64_p2.magic_bias; 30 const float vminus_log2e = params->scalar_rr2_lut64_p2.minus_log2e; 31 const uint32_t vindex_mask = UINT32_C(0x3F); 32 const float vln2_hi = params->scalar_rr2_lut64_p2.ln2_hi; 33 const float vln2_lo = params->scalar_rr2_lut64_p2.ln2_lo; 34 const float vc2 = params->scalar_rr2_lut64_p2.c2; 35 const float vone = params->scalar_rr2_lut64_p2.one; 36 const float vdenorm_cutoff = params->scalar_rr2_lut64_p2.denorm_cutoff; 37 38 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 39 const float vx0 = x[0]; 40 const float vx1 = x[1]; 41 const float vx2 = x[2]; 42 const float vx3 = x[3]; 43 x += 4; 44 45 const float vz0 = fabsf(vx0); 46 const float vz1 = fabsf(vx1); 47 const float vz2 = fabsf(vx2); 48 const float vz3 = fabsf(vx3); 49 50 float vn0 = vz0 * vminus_log2e + vmagic_bias; 51 float vn1 = vz1 * vminus_log2e + vmagic_bias; 52 float vn2 = vz2 * vminus_log2e + vmagic_bias; 53 float vn3 = vz3 * vminus_log2e + vmagic_bias; 54 55 const uint32_t ve0 = float_as_uint32(vn0) << 17; 56 const uint32_t ve1 = float_as_uint32(vn1) << 17; 57 const uint32_t ve2 = float_as_uint32(vn2) << 17; 58 const uint32_t ve3 = float_as_uint32(vn3) << 17; 59 60 const uint32_t vidx0 = float_as_uint32(vn0) & vindex_mask; 61 const float vs0 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx0] + ve0); 62 const uint32_t vidx1 = float_as_uint32(vn1) & vindex_mask; 63 const float vs1 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx1] + ve1); 64 const uint32_t vidx2 = float_as_uint32(vn2) & vindex_mask; 65 const float vs2 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx2] + ve2); 66 const uint32_t vidx3 = float_as_uint32(vn3) & vindex_mask; 67 const float vs3 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx3] + ve3); 68 69 vn0 -= vmagic_bias; 70 vn1 -= vmagic_bias; 71 vn2 -= vmagic_bias; 72 vn3 -= vmagic_bias; 73 74 float vt0 = vn0 * vln2_hi + vz0; 75 float vt1 = vn1 * vln2_hi + vz1; 76 float vt2 = vn2 * vln2_hi + vz2; 77 float vt3 = vn3 * vln2_hi + vz3; 78 79 vt0 = vn0 * vln2_lo + vt0; 80 vt1 = vn1 * vln2_lo + vt1; 81 vt2 = vn2 * vln2_lo + vt2; 82 vt3 = vn3 * vln2_lo + vt3; 83 84 float vp0 = vt0 * vc2; 85 float vp1 = vt1 * vc2; 86 float vp2 = vt2 * vc2; 87 float vp3 = vt3 * vc2; 88 89 vp0 = vt0 - vp0 * vt0; 90 vp1 = vt1 - vp1 * vt1; 91 vp2 = vt2 - vp2 * vt2; 92 vp3 = vt3 - vp3 * vt3; 93 94 const float vy0 = vs0 - vs0 * vp0; 95 const float vy1 = vs1 - vs1 * vp1; 96 const float vy2 = vs2 - vs2 * vp2; 97 const float vy3 = vs3 - vs3 * vp3; 98 99 const float vd0 = vy0 + vone; 100 const float vd1 = vy1 + vone; 101 const float vd2 = vy2 + vone; 102 const float vd3 = vy3 + vone; 103 104 float vf0 = vy0 / vd0; 105 float vf1 = vy1 / vd1; 106 float vf2 = vy2 / vd2; 107 float vf3 = vy3 / vd3; 108 109 if XNN_UNPREDICTABLE(vz0 > vdenorm_cutoff) { 110 vf0 = 0.0f; 111 } 112 if XNN_UNPREDICTABLE(vz1 > vdenorm_cutoff) { 113 vf1 = 0.0f; 114 } 115 if XNN_UNPREDICTABLE(vz2 > vdenorm_cutoff) { 116 vf2 = 0.0f; 117 } 118 if XNN_UNPREDICTABLE(vz3 > vdenorm_cutoff) { 119 vf3 = 0.0f; 120 } 121 122 if XNN_UNPREDICTABLE(vx0 > 0.0f) { 123 vf0 = vone - vf0; 124 } 125 if XNN_UNPREDICTABLE(vx1 > 0.0f) { 126 vf1 = vone - vf1; 127 } 128 if XNN_UNPREDICTABLE(vx2 > 0.0f) { 129 vf2 = vone - vf2; 130 } 131 if XNN_UNPREDICTABLE(vx3 > 0.0f) { 132 vf3 = vone - vf3; 133 } 134 135 y[0] = vf0; 136 y[1] = vf1; 137 y[2] = vf2; 138 y[3] = vf3; 139 y += 4; 140 } 141 if XNN_UNLIKELY(n != 0) { 142 do { 143 const float vx = *x++; 144 145 const float vz = fabsf(vx); 146 147 float vn = vz * vminus_log2e + vmagic_bias; 148 const uint32_t ve = float_as_uint32(vn) << 17; 149 const uint32_t vidx = float_as_uint32(vn) & vindex_mask; 150 const float vs = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx] + ve); 151 vn -= vmagic_bias; 152 153 float vt = vn * vln2_hi + vz; 154 vt = vn * vln2_lo + vt; 155 156 float vp = vt * vc2; 157 vp = vt - vp * vt; 158 159 const float vy = vs - vs * vp; 160 const float vd = vy + vone; 161 162 float vf = vy / vd; 163 if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) { 164 vf = 0.0f; 165 } 166 if XNN_UNPREDICTABLE(vx > 0.0f) { 167 vf = vone - vf; 168 } 169 170 *y++ = vf; 171 172 n -= sizeof(float); 173 } while (n != 0); 174 } 175 } 176