xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x20.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vsigmoid/neon-lut64-p2.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
18*4bdc9457SAndroid Build Coastguard Worker extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20(
21*4bdc9457SAndroid Build Coastguard Worker     size_t n,
22*4bdc9457SAndroid Build Coastguard Worker     const float* x,
23*4bdc9457SAndroid Build Coastguard Worker     float* y,
24*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25*4bdc9457SAndroid Build Coastguard Worker {
26*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
27*4bdc9457SAndroid Build Coastguard Worker 
28*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neon_rr2_lut64_p2.magic_bias);
29*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neon_rr2_lut64_p2.minus_log2e);
30*4bdc9457SAndroid Build Coastguard Worker   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vln2_hi = vld1q_dup_f32(&params->neon_rr2_lut64_p2.ln2_hi);
32*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vln2_lo = vld1q_dup_f32(&params->neon_rr2_lut64_p2.ln2_lo);
33*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vc2 = vld1q_dup_f32(&params->neon_rr2_lut64_p2.c2);
34*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vone = vmovq_n_f32(1.0f);
35*4bdc9457SAndroid Build Coastguard Worker   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neon_rr2_lut64_p2.denorm_cutoff);
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
38*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vx89AB = vld1q_f32(x); x += 4;
41*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
42*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
43*4bdc9457SAndroid Build Coastguard Worker 
44*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vz0123 = vabsq_f32(vx0123);
45*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vz4567 = vabsq_f32(vx4567);
46*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vz89AB = vabsq_f32(vx89AB);
47*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
48*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
49*4bdc9457SAndroid Build Coastguard Worker 
50*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
51*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
52*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
53*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
54*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
55*4bdc9457SAndroid Build Coastguard Worker 
56*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
57*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
58*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
59*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
60*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t veGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 17);
61*4bdc9457SAndroid Build Coastguard Worker 
62*4bdc9457SAndroid Build Coastguard Worker     // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
63*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
64*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
65*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
66*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
67*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask));
68*4bdc9457SAndroid Build Coastguard Worker 
69*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
70*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
71*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
72*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
73*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
74*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
75*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
76*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
77*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
78*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
79*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
80*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
81*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
82*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
83*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
84*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
85*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
86*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
87*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vlGH = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxGH]);
88*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]);
89*4bdc9457SAndroid Build Coastguard Worker 
90*4bdc9457SAndroid Build Coastguard Worker     vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
91*4bdc9457SAndroid Build Coastguard Worker     vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
92*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
93*4bdc9457SAndroid Build Coastguard Worker     vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
94*4bdc9457SAndroid Build Coastguard Worker     vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
95*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
96*4bdc9457SAndroid Build Coastguard Worker     vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
97*4bdc9457SAndroid Build Coastguard Worker     vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
98*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
99*4bdc9457SAndroid Build Coastguard Worker     vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
100*4bdc9457SAndroid Build Coastguard Worker     vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
101*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
102*4bdc9457SAndroid Build Coastguard Worker     vlGH = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxGH >> 32)], vlGH, 1);
103*4bdc9457SAndroid Build Coastguard Worker     vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1);
104*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ);
105*4bdc9457SAndroid Build Coastguard Worker 
106*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
107*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
108*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
109*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
110*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ));
111*4bdc9457SAndroid Build Coastguard Worker 
112*4bdc9457SAndroid Build Coastguard Worker     vn0123 = vsubq_f32(vn0123, vmagic_bias);
113*4bdc9457SAndroid Build Coastguard Worker     vn4567 = vsubq_f32(vn4567, vmagic_bias);
114*4bdc9457SAndroid Build Coastguard Worker     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
115*4bdc9457SAndroid Build Coastguard Worker     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
116*4bdc9457SAndroid Build Coastguard Worker     vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
117*4bdc9457SAndroid Build Coastguard Worker 
118*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
119*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
120*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
121*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
122*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);
123*4bdc9457SAndroid Build Coastguard Worker 
124*4bdc9457SAndroid Build Coastguard Worker     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
125*4bdc9457SAndroid Build Coastguard Worker     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
126*4bdc9457SAndroid Build Coastguard Worker     vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
127*4bdc9457SAndroid Build Coastguard Worker     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
128*4bdc9457SAndroid Build Coastguard Worker     vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);
129*4bdc9457SAndroid Build Coastguard Worker 
130*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
131*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
132*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
133*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
134*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vpGHIJ = vmulq_f32(vtGHIJ, vc2);
135*4bdc9457SAndroid Build Coastguard Worker 
136*4bdc9457SAndroid Build Coastguard Worker     vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
137*4bdc9457SAndroid Build Coastguard Worker     vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
138*4bdc9457SAndroid Build Coastguard Worker     vp89AB = vmlsq_f32(vt89AB, vp89AB, vt89AB);
139*4bdc9457SAndroid Build Coastguard Worker     vpCDEF = vmlsq_f32(vtCDEF, vpCDEF, vtCDEF);
140*4bdc9457SAndroid Build Coastguard Worker     vpGHIJ = vmlsq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
141*4bdc9457SAndroid Build Coastguard Worker 
142*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
143*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
144*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB);
145*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF);
146*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vyGHIJ = vmlsq_f32(vsGHIJ, vsGHIJ, vpGHIJ);
147*4bdc9457SAndroid Build Coastguard Worker 
148*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
149*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
150*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
151*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
152*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vdGHIJ = vaddq_f32(vyGHIJ, vone);
153*4bdc9457SAndroid Build Coastguard Worker 
154*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vr0123 = vrecpeq_f32(vd0123);
155*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vr4567 = vrecpeq_f32(vd4567);
156*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vr89AB = vrecpeq_f32(vd89AB);
157*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
158*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
159*4bdc9457SAndroid Build Coastguard Worker 
160*4bdc9457SAndroid Build Coastguard Worker     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
161*4bdc9457SAndroid Build Coastguard Worker     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
162*4bdc9457SAndroid Build Coastguard Worker     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
163*4bdc9457SAndroid Build Coastguard Worker     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
164*4bdc9457SAndroid Build Coastguard Worker     vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
165*4bdc9457SAndroid Build Coastguard Worker 
166*4bdc9457SAndroid Build Coastguard Worker     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
167*4bdc9457SAndroid Build Coastguard Worker     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
168*4bdc9457SAndroid Build Coastguard Worker     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
169*4bdc9457SAndroid Build Coastguard Worker     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
170*4bdc9457SAndroid Build Coastguard Worker     vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
171*4bdc9457SAndroid Build Coastguard Worker 
172*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
173*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
174*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
175*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
176*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vfGHIJ = vmulq_f32(vyGHIJ, vrGHIJ);
177*4bdc9457SAndroid Build Coastguard Worker 
178*4bdc9457SAndroid Build Coastguard Worker     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
179*4bdc9457SAndroid Build Coastguard Worker     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
180*4bdc9457SAndroid Build Coastguard Worker     vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
181*4bdc9457SAndroid Build Coastguard Worker     vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
182*4bdc9457SAndroid Build Coastguard Worker     vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
183*4bdc9457SAndroid Build Coastguard Worker 
184*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
185*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
186*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
187*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
188*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
189*4bdc9457SAndroid Build Coastguard Worker 
190*4bdc9457SAndroid Build Coastguard Worker     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
191*4bdc9457SAndroid Build Coastguard Worker     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
192*4bdc9457SAndroid Build Coastguard Worker     vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
193*4bdc9457SAndroid Build Coastguard Worker     vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
194*4bdc9457SAndroid Build Coastguard Worker     vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
195*4bdc9457SAndroid Build Coastguard Worker 
196*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vf0123); y += 4;
197*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vf4567); y += 4;
198*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vf89AB); y += 4;
199*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vfCDEF); y += 4;
200*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vfGHIJ); y += 4;
201*4bdc9457SAndroid Build Coastguard Worker   }
202*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
203*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vx = vld1q_f32(x); x += 4;
204*4bdc9457SAndroid Build Coastguard Worker 
205*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vz = vabsq_f32(vx);
206*4bdc9457SAndroid Build Coastguard Worker 
207*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
208*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
209*4bdc9457SAndroid Build Coastguard Worker 
210*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
211*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
212*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
213*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
214*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
215*4bdc9457SAndroid Build Coastguard Worker     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
216*4bdc9457SAndroid Build Coastguard Worker     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
217*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
218*4bdc9457SAndroid Build Coastguard Worker 
219*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
220*4bdc9457SAndroid Build Coastguard Worker     vn = vsubq_f32(vn, vmagic_bias);
221*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
222*4bdc9457SAndroid Build Coastguard Worker     vt = vmlaq_f32(vt, vn, vln2_lo);
223*4bdc9457SAndroid Build Coastguard Worker 
224*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vp = vmulq_f32(vt, vc2);
225*4bdc9457SAndroid Build Coastguard Worker     vp = vmlsq_f32(vt, vp, vt);
226*4bdc9457SAndroid Build Coastguard Worker 
227*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vy = vmlsq_f32(vs, vs, vp);
228*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vd = vaddq_f32(vy, vone);
229*4bdc9457SAndroid Build Coastguard Worker 
230*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vr = vrecpeq_f32(vd);
231*4bdc9457SAndroid Build Coastguard Worker     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
232*4bdc9457SAndroid Build Coastguard Worker     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
233*4bdc9457SAndroid Build Coastguard Worker 
234*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vf = vmulq_f32(vy, vr);
235*4bdc9457SAndroid Build Coastguard Worker     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
236*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
237*4bdc9457SAndroid Build Coastguard Worker     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
238*4bdc9457SAndroid Build Coastguard Worker 
239*4bdc9457SAndroid Build Coastguard Worker     vst1q_f32(y, vf); y += 4;
240*4bdc9457SAndroid Build Coastguard Worker   }
241*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
242*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vx = vld1q_f32(x);
243*4bdc9457SAndroid Build Coastguard Worker 
244*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vz = vabsq_f32(vx);
245*4bdc9457SAndroid Build Coastguard Worker 
246*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
247*4bdc9457SAndroid Build Coastguard Worker     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
248*4bdc9457SAndroid Build Coastguard Worker 
249*4bdc9457SAndroid Build Coastguard Worker     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
250*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
251*4bdc9457SAndroid Build Coastguard Worker     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
252*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
253*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
254*4bdc9457SAndroid Build Coastguard Worker     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
255*4bdc9457SAndroid Build Coastguard Worker     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
256*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
257*4bdc9457SAndroid Build Coastguard Worker 
258*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
259*4bdc9457SAndroid Build Coastguard Worker     vn = vsubq_f32(vn, vmagic_bias);
260*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
261*4bdc9457SAndroid Build Coastguard Worker     vt = vmlaq_f32(vt, vn, vln2_lo);
262*4bdc9457SAndroid Build Coastguard Worker 
263*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vp = vmulq_f32(vt, vc2);
264*4bdc9457SAndroid Build Coastguard Worker     vp = vmlsq_f32(vt, vp, vt);
265*4bdc9457SAndroid Build Coastguard Worker 
266*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vy = vmlsq_f32(vs, vs, vp);
267*4bdc9457SAndroid Build Coastguard Worker     const float32x4_t vd = vaddq_f32(vy, vone);
268*4bdc9457SAndroid Build Coastguard Worker 
269*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vr = vrecpeq_f32(vd);
270*4bdc9457SAndroid Build Coastguard Worker     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
271*4bdc9457SAndroid Build Coastguard Worker     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
272*4bdc9457SAndroid Build Coastguard Worker 
273*4bdc9457SAndroid Build Coastguard Worker     float32x4_t vf = vmulq_f32(vy, vr);
274*4bdc9457SAndroid Build Coastguard Worker     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
275*4bdc9457SAndroid Build Coastguard Worker     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
276*4bdc9457SAndroid Build Coastguard Worker     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
277*4bdc9457SAndroid Build Coastguard Worker 
278*4bdc9457SAndroid Build Coastguard Worker     float32x2_t vf_lo = vget_low_f32(vf);
279*4bdc9457SAndroid Build Coastguard Worker     if (n & (2 * sizeof(float))) {
280*4bdc9457SAndroid Build Coastguard Worker       vst1_f32(y, vf_lo); y += 2;
281*4bdc9457SAndroid Build Coastguard Worker       vf_lo = vget_high_f32(vf);
282*4bdc9457SAndroid Build Coastguard Worker     }
283*4bdc9457SAndroid Build Coastguard Worker     if (n & (1 * sizeof(float))) {
284*4bdc9457SAndroid Build Coastguard Worker       vst1_lane_f32(y, vf_lo, 0);
285*4bdc9457SAndroid Build Coastguard Worker     }
286*4bdc9457SAndroid Build Coastguard Worker   }
287*4bdc9457SAndroid Build Coastguard Worker }
288