xref: /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/vsigmoid-neonfp16arith-rr2-p2-div-x32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vsigmoid/neonfp16arith.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 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 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x32(size_t batch,const void * input,void * output,const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x32(
19     size_t batch,
20     const void* input,
21     void* output,
22     const union xnn_f16_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(batch % sizeof(__fp16) == 0);
25 
26   const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.magic_bias));
27   const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_log2e));
28   const float16x8_t vln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.ln2_hi));
29   const float16x8_t vln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.ln2_lo));
30   const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c2));
31   const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c1));
32   const float16x8_t vone = vmovq_n_f16(1.0f);
33   const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.denorm_cutoff));
34 
35   const __fp16* i = (const __fp16*) input;
36   __fp16* o = (__fp16*) output;
37   for (; batch >= 32 * sizeof(__fp16); batch -= 32 * sizeof(__fp16)) {
38     const float16x8_t vx0 = vld1q_f16(i); i += 8;
39     const float16x8_t vx1 = vld1q_f16(i); i += 8;
40     const float16x8_t vx2 = vld1q_f16(i); i += 8;
41     const float16x8_t vx3 = vld1q_f16(i); i += 8;
42 
43     const float16x8_t vz0 = vabsq_f16(vx0);
44     const float16x8_t vz1 = vabsq_f16(vx1);
45     const float16x8_t vz2 = vabsq_f16(vx2);
46     const float16x8_t vz3 = vabsq_f16(vx3);
47 
48     float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
49     float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
50     float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
51     float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
52 
53     const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
54     const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
55     const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
56     const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
57 
58     vn0 = vsubq_f16(vn0, vmagic_bias);
59     vn1 = vsubq_f16(vn1, vmagic_bias);
60     vn2 = vsubq_f16(vn2, vmagic_bias);
61     vn3 = vsubq_f16(vn3, vmagic_bias);
62 
63     float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2_hi);
64     float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2_hi);
65     float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2_hi);
66     float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2_hi);
67 
68     vt0 = vfmaq_f16(vt0, vn0, vln2_lo);
69     vt1 = vfmaq_f16(vt1, vn1, vln2_lo);
70     vt2 = vfmaq_f16(vt2, vn2, vln2_lo);
71     vt3 = vfmaq_f16(vt3, vn3, vln2_lo);
72 
73     const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
74     const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
75     const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
76     const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
77 
78     vt0 = vmulq_f16(vt0, vs0);
79     vt1 = vmulq_f16(vt1, vs1);
80     vt2 = vmulq_f16(vt2, vs2);
81     vt3 = vmulq_f16(vt3, vs3);
82 
83     const float16x8_t ve0 = vfmaq_f16(vs0, vp0, vt0);
84     const float16x8_t ve1 = vfmaq_f16(vs1, vp1, vt1);
85     const float16x8_t ve2 = vfmaq_f16(vs2, vp2, vt2);
86     const float16x8_t ve3 = vfmaq_f16(vs3, vp3, vt3);
87 
88     const float16x8_t vd0 = vaddq_f16(ve0, vone);
89     const float16x8_t vd1 = vaddq_f16(ve1, vone);
90     const float16x8_t vd2 = vaddq_f16(ve2, vone);
91     const float16x8_t vd3 = vaddq_f16(ve3, vone);
92 
93     float16x8_t vf0 = vdivq_f16(ve0, vd0);
94     float16x8_t vf1 = vdivq_f16(ve1, vd1);
95     float16x8_t vf2 = vdivq_f16(ve2, vd2);
96     float16x8_t vf3 = vdivq_f16(ve3, vd3);
97 
98     vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vcagtq_f16(vx0, vdenorm_cutoff)));
99     vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vcagtq_f16(vx1, vdenorm_cutoff)));
100     vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vcagtq_f16(vx2, vdenorm_cutoff)));
101     vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vcagtq_f16(vx3, vdenorm_cutoff)));
102 
103     const uint16x8_t vm0 = vcltq_f16(vx0, vmovq_n_f16(0.0f));
104     const uint16x8_t vm1 = vcltq_f16(vx1, vmovq_n_f16(0.0f));
105     const uint16x8_t vm2 = vcltq_f16(vx2, vmovq_n_f16(0.0f));
106     const uint16x8_t vm3 = vcltq_f16(vx3, vmovq_n_f16(0.0f));
107 
108     vf0 = vbslq_f16(vm0, vf0, vsubq_f16(vone, vf0));
109     vf1 = vbslq_f16(vm1, vf1, vsubq_f16(vone, vf1));
110     vf2 = vbslq_f16(vm2, vf2, vsubq_f16(vone, vf2));
111     vf3 = vbslq_f16(vm3, vf3, vsubq_f16(vone, vf3));
112 
113     vst1q_f16(o, vf0); o += 8;
114     vst1q_f16(o, vf1); o += 8;
115     vst1q_f16(o, vf2); o += 8;
116     vst1q_f16(o, vf3); o += 8;
117   }
118   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
119     const float16x8_t vx = vld1q_f16(i); i += 8;
120 
121     const float16x8_t vz = vabsq_f16(vx);
122 
123     float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
124     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
125     vn = vsubq_f16(vn, vmagic_bias);
126 
127     float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
128     vt = vfmaq_f16(vt, vn, vln2_lo);
129 
130     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
131     vt = vmulq_f16(vt, vs);
132     const float16x8_t ve = vfmaq_f16(vs, vp, vt);
133     const float16x8_t vd = vaddq_f16(ve, vone);
134 
135     float16x8_t vf = vdivq_f16(ve, vd);
136     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
137     const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
138     vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
139 
140     vst1q_f16(o, vf); o += 8;
141   }
142   if XNN_UNLIKELY(batch != 0) {
143     const float16x8_t vx = vld1q_f16(i);
144 
145     const float16x8_t vz = vabsq_f16(vx);
146 
147     float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
148     const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
149     vn = vsubq_f16(vn, vmagic_bias);
150 
151     float16x8_t vt = vfmaq_f16(vz, vn, vln2_hi);
152     vt = vfmaq_f16(vt, vn, vln2_lo);
153 
154     const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
155     vt = vmulq_f16(vt, vs);
156     const float16x8_t ve = vfmaq_f16(vs, vp, vt);
157     const float16x8_t vd = vaddq_f16(ve, vone);
158 
159     float16x8_t vf = vdivq_f16(ve, vd);
160     vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vcagtq_f16(vx, vdenorm_cutoff)));
161     const uint16x8_t vm = vcltq_f16(vx, vmovq_n_f16(0.0f));
162     vf = vbslq_f16(vm, vf, vsubq_f16(vone, vf));
163 
164     float16x4_t vf_lo = vget_low_f16(vf);
165     if (batch & (4 * sizeof(__fp16))) {
166       vst1_f16(o, vf_lo); o += 4;
167       vf_lo = vget_high_f16(vf);
168     }
169     if (batch & (2 * sizeof(__fp16))) {
170       vst1_f16(o, vf_lo); o += 2;
171       vf_lo = vext_f16(vf_lo, vf_lo, 2);
172     }
173     if (batch & (1 * sizeof(__fp16))) {
174       vst1_lane_f16(o, vf_lo, 0);
175     }
176   }
177 }
178