xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-lut64-p2-div-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__neonfma_rr1_lut64_p2_div_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__neonfma_rr1_lut64_p2_div_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->neonfma_rr1_lut64_p2.magic_bias);
29   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.minus_log2e);
30   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31   const float32x4_t vln2 = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.ln2);
32   const float32x4_t vc2 = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.c2);
33   const float32x4_t vone = vmovq_n_f32(1.0f);
34   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.denorm_cutoff);
35 
36   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
37     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
38     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
39 
40     const float32x4_t vz0123 = vabsq_f32(vx0123);
41     const float32x4_t vz4567 = vabsq_f32(vx4567);
42 
43     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
44     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
45 
46     const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
47     const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
48 
49     // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
50     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
51     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
52 
53     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
54     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
55     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
56     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
57     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
58     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
59     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
60     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
61 
62     vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
63     vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
64     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
65     vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
66     vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
67     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
68 
69     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
70     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
71 
72     vn0123 = vsubq_f32(vn0123, vmagic_bias);
73     vn4567 = vsubq_f32(vn4567, vmagic_bias);
74 
75     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
76     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
77 
78     float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
79     float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
80 
81     vp0123 = vfmsq_f32(vt0123, vp0123, vt0123);
82     vp4567 = vfmsq_f32(vt4567, vp4567, vt4567);
83 
84     const float32x4_t vy0123 = vfmsq_f32(vs0123, vs0123, vp0123);
85     const float32x4_t vy4567 = vfmsq_f32(vs4567, vs4567, vp4567);
86 
87     const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
88     const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
89 
90     float32x4_t vf0123 = vdivq_f32(vy0123, vd0123);
91     float32x4_t vf4567 = vdivq_f32(vy4567, vd4567);
92 
93     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
94     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
95 
96     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
97     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
98 
99     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
100     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
101 
102     vst1q_f32(y, vf0123); y += 4;
103     vst1q_f32(y, vf4567); y += 4;
104   }
105   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
106     const float32x4_t vx = vld1q_f32(x); x += 4;
107 
108     const float32x4_t vz = vabsq_f32(vx);
109 
110     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
111     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
112 
113     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
114     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
115     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
116     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
117     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
118     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
119     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
120     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
121 
122     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
123     vn = vsubq_f32(vn, vmagic_bias);
124     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
125 
126     float32x4_t vp = vmulq_f32(vt, vc2);
127     vp = vfmsq_f32(vt, vp, vt);
128 
129     const float32x4_t vy = vfmsq_f32(vs, vs, vp);
130     const float32x4_t vd = vaddq_f32(vy, vone);
131 
132     float32x4_t vf = vdivq_f32(vy, vd);
133     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
134     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
135     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
136 
137     vst1q_f32(y, vf); y += 4;
138   }
139   if XNN_UNLIKELY(n != 0) {
140     const float32x4_t vx = vld1q_f32(x);
141 
142     const float32x4_t vz = vabsq_f32(vx);
143 
144     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
145     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
146 
147     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
148     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
149     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
150     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
151     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
152     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
153     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
154     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
155 
156     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
157     vn = vsubq_f32(vn, vmagic_bias);
158     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
159 
160     float32x4_t vp = vmulq_f32(vt, vc2);
161     vp = vfmsq_f32(vt, vp, vt);
162 
163     const float32x4_t vy = vfmsq_f32(vs, vs, vp);
164     const float32x4_t vd = vaddq_f32(vy, vone);
165 
166     float32x4_t vf = vdivq_f32(vy, vd);
167     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
168     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
169     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
170 
171     float32x2_t vf_lo = vget_low_f32(vf);
172     if (n & (2 * sizeof(float))) {
173       vst1_f32(y, vf_lo); y += 2;
174       vf_lo = vget_high_f32(vf);
175     }
176     if (n & (1 * sizeof(float))) {
177       vst1_lane_f32(y, vf_lo, 0);
178     }
179   }
180 }
181