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_x2(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_x2( 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 >= 2 * sizeof(float); n -= 2 * sizeof(float)) { 39 const float vx0 = x[0]; 40 const float vx1 = x[1]; 41 x += 2; 42 43 const float vz0 = fabsf(vx0); 44 const float vz1 = fabsf(vx1); 45 46 float vn0 = vz0 * vminus_log2e + vmagic_bias; 47 float vn1 = vz1 * vminus_log2e + vmagic_bias; 48 49 const uint32_t ve0 = float_as_uint32(vn0) << 17; 50 const uint32_t ve1 = float_as_uint32(vn1) << 17; 51 52 const uint32_t vidx0 = float_as_uint32(vn0) & vindex_mask; 53 const float vs0 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx0] + ve0); 54 const uint32_t vidx1 = float_as_uint32(vn1) & vindex_mask; 55 const float vs1 = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx1] + ve1); 56 57 vn0 -= vmagic_bias; 58 vn1 -= vmagic_bias; 59 60 float vt0 = vn0 * vln2_hi + vz0; 61 float vt1 = vn1 * vln2_hi + vz1; 62 63 vt0 = vn0 * vln2_lo + vt0; 64 vt1 = vn1 * vln2_lo + vt1; 65 66 float vp0 = vt0 * vc2; 67 float vp1 = vt1 * vc2; 68 69 vp0 = vt0 - vp0 * vt0; 70 vp1 = vt1 - vp1 * vt1; 71 72 const float vy0 = vs0 - vs0 * vp0; 73 const float vy1 = vs1 - vs1 * vp1; 74 75 const float vd0 = vy0 + vone; 76 const float vd1 = vy1 + vone; 77 78 float vf0 = vy0 / vd0; 79 float vf1 = vy1 / vd1; 80 81 if XNN_UNPREDICTABLE(vz0 > vdenorm_cutoff) { 82 vf0 = 0.0f; 83 } 84 if XNN_UNPREDICTABLE(vz1 > vdenorm_cutoff) { 85 vf1 = 0.0f; 86 } 87 88 if XNN_UNPREDICTABLE(vx0 > 0.0f) { 89 vf0 = vone - vf0; 90 } 91 if XNN_UNPREDICTABLE(vx1 > 0.0f) { 92 vf1 = vone - vf1; 93 } 94 95 y[0] = vf0; 96 y[1] = vf1; 97 y += 2; 98 } 99 if XNN_UNLIKELY(n != 0) { 100 const float vx = *x; 101 102 const float vz = fabsf(vx); 103 104 float vn = vz * vminus_log2e + vmagic_bias; 105 const uint32_t ve = float_as_uint32(vn) << 17; 106 const uint32_t vidx = float_as_uint32(vn) & vindex_mask; 107 const float vs = uint32_as_float(xnn_table_exp2minus_k_over_64[vidx] + ve); 108 vn -= vmagic_bias; 109 110 float vt = vn * vln2_hi + vz; 111 vt = vn * vln2_lo + vt; 112 113 float vp = vt * vc2; 114 vp = vt - vp * vt; 115 116 const float vy = vs - vs * vp; 117 const float vd = vy + vone; 118 119 float vf = vy / vd; 120 if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) { 121 vf = 0.0f; 122 } 123 if XNN_UNPREDICTABLE(vx > 0.0f) { 124 vf = vone - vf; 125 } 126 127 *y = vf; 128 } 129 } 130