xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neon-rr2-lut64-p2-nr2recps-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/neon-lut64-p2.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 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19 
xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8(
21     size_t n,
22     const float* x,
23     float* y,
24     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26   assert(n % sizeof(float) == 0);
27 
28   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neon_rr2_lut64_p2.magic_bias);
29   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neon_rr2_lut64_p2.minus_log2e);
30   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31   const float32x4_t vln2_hi = vld1q_dup_f32(&params->neon_rr2_lut64_p2.ln2_hi);
32   const float32x4_t vln2_lo = vld1q_dup_f32(&params->neon_rr2_lut64_p2.ln2_lo);
33   const float32x4_t vc2 = vld1q_dup_f32(&params->neon_rr2_lut64_p2.c2);
34   const float32x4_t vone = vmovq_n_f32(1.0f);
35   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neon_rr2_lut64_p2.denorm_cutoff);
36 
37   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
38     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40 
41     const float32x4_t vz0123 = vabsq_f32(vx0123);
42     const float32x4_t vz4567 = vabsq_f32(vx4567);
43 
44     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
45     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
46 
47     const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
48     const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
49 
50     // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
51     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
52     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
53 
54     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
55     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
56     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
57     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
58     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
59     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
60     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
61     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
62 
63     vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
64     vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
65     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
66     vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
67     vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
68     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
69 
70     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
71     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
72 
73     vn0123 = vsubq_f32(vn0123, vmagic_bias);
74     vn4567 = vsubq_f32(vn4567, vmagic_bias);
75 
76     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
77     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
78 
79     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
80     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
81 
82     float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
83     float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
84 
85     vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
86     vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
87 
88     const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
89     const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
90 
91     const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
92     const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
93 
94     float32x4_t vr0123 = vrecpeq_f32(vd0123);
95     float32x4_t vr4567 = vrecpeq_f32(vd4567);
96 
97     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
98     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
99 
100     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
101     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
102 
103     float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
104     float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
105 
106     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
107     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
108 
109     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
110     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
111 
112     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
113     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
114 
115     vst1q_f32(y, vf0123); y += 4;
116     vst1q_f32(y, vf4567); y += 4;
117   }
118   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
119     const float32x4_t vx = vld1q_f32(x); x += 4;
120 
121     const float32x4_t vz = vabsq_f32(vx);
122 
123     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
124     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
125 
126     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
127     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
128     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
129     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
130     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
131     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
132     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
133     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
134 
135     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
136     vn = vsubq_f32(vn, vmagic_bias);
137     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
138     vt = vmlaq_f32(vt, vn, vln2_lo);
139 
140     float32x4_t vp = vmulq_f32(vt, vc2);
141     vp = vmlsq_f32(vt, vp, vt);
142 
143     const float32x4_t vy = vmlsq_f32(vs, vs, vp);
144     const float32x4_t vd = vaddq_f32(vy, vone);
145 
146     float32x4_t vr = vrecpeq_f32(vd);
147     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
148     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
149 
150     float32x4_t vf = vmulq_f32(vy, vr);
151     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
152     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
153     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
154 
155     vst1q_f32(y, vf); y += 4;
156   }
157   if XNN_UNLIKELY(n != 0) {
158     const float32x4_t vx = vld1q_f32(x);
159 
160     const float32x4_t vz = vabsq_f32(vx);
161 
162     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
163     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
164 
165     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
166     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
167     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
168     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
169     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
170     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
171     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
172     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
173 
174     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
175     vn = vsubq_f32(vn, vmagic_bias);
176     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
177     vt = vmlaq_f32(vt, vn, vln2_lo);
178 
179     float32x4_t vp = vmulq_f32(vt, vc2);
180     vp = vmlsq_f32(vt, vp, vt);
181 
182     const float32x4_t vy = vmlsq_f32(vs, vs, vp);
183     const float32x4_t vd = vaddq_f32(vy, vone);
184 
185     float32x4_t vr = vrecpeq_f32(vd);
186     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
187     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
188 
189     float32x4_t vf = vmulq_f32(vy, vr);
190     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
191     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
192     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
193 
194     float32x2_t vf_lo = vget_low_f32(vf);
195     if (n & (2 * sizeof(float))) {
196       vst1_f32(y, vf_lo); y += 2;
197       vf_lo = vget_high_f32(vf);
198     }
199     if (n & (1 * sizeof(float))) {
200       vst1_lane_f32(y, vf_lo, 0);
201     }
202   }
203 }
204