• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/neon-lut2048-p1.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_2048[2048];
19 
xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16(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_lut2048_p1_nr2recps_x16(
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_lut2048_p1.magic_bias);
29   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neonfma_rr1_lut2048_p1.minus_log2e);
30   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
31   const float32x4_t vln2 = vld1q_dup_f32(&params->neonfma_rr1_lut2048_p1.ln2);
32   const float32x4_t vc1 = vld1q_dup_f32(&params->neonfma_rr1_lut2048_p1.c1);
33   const float32x4_t vone = vmovq_n_f32(1.0f);
34   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neonfma_rr1_lut2048_p1.denorm_cutoff);
35 
36   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
37     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
38     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
39     const float32x4_t vx89AB = vld1q_f32(x); x += 4;
40     const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
41 
42     const float32x4_t vz0123 = vabsq_f32(vx0123);
43     const float32x4_t vz4567 = vabsq_f32(vx4567);
44     const float32x4_t vz89AB = vabsq_f32(vx89AB);
45     const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
46 
47     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
48     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
49     float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
50     float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
51 
52     const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 12);
53     const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 12);
54     const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 12);
55     const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 12);
56 
57     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
58     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
59     const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
60     const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
61 
62     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
63     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
64     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx01]);
65     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx23]);
66     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
67     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
68     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx45]);
69     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx67]);
70     const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
71     const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
72     float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx89]);
73     float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxAB]);
74     const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
75     const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
76     float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxCD]);
77     float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxEF]);
78 
79     vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx01 >> 32)], vl01, 1);
80     vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx23 >> 32)], vl23, 1);
81     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
82     vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx45 >> 32)], vl45, 1);
83     vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx67 >> 32)], vl67, 1);
84     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
85     vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx89 >> 32)], vl89, 1);
86     vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxAB >> 32)], vlAB, 1);
87     const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
88     vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxCD >> 32)], vlCD, 1);
89     vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxEF >> 32)], vlEF, 1);
90     const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
91 
92     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
93     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
94     const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
95     const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
96 
97     vn0123 = vsubq_f32(vn0123, vmagic_bias);
98     vn4567 = vsubq_f32(vn4567, vmagic_bias);
99     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
100     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
101 
102     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
103     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
104     float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
105     float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
106 
107     const float32x4_t vp0123 = vmulq_f32(vt0123, vc1);
108     const float32x4_t vp4567 = vmulq_f32(vt4567, vc1);
109     const float32x4_t vp89AB = vmulq_f32(vt89AB, vc1);
110     const float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc1);
111 
112     const float32x4_t vy0123 = vfmaq_f32(vs0123, vs0123, vp0123);
113     const float32x4_t vy4567 = vfmaq_f32(vs4567, vs4567, vp4567);
114     const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB);
115     const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF);
116 
117     const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
118     const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
119     const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
120     const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
121 
122     float32x4_t vr0123 = vrecpeq_f32(vd0123);
123     float32x4_t vr4567 = vrecpeq_f32(vd4567);
124     float32x4_t vr89AB = vrecpeq_f32(vd89AB);
125     float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
126 
127     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
128     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
129     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
130     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
131 
132     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
133     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
134     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
135     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
136 
137     float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
138     float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
139     float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
140     float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
141 
142     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
143     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
144     vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
145     vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
146 
147     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
148     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
149     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
150     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
151 
152     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
153     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
154     vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
155     vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
156 
157     vst1q_f32(y, vf0123); y += 4;
158     vst1q_f32(y, vf4567); y += 4;
159     vst1q_f32(y, vf89AB); y += 4;
160     vst1q_f32(y, vfCDEF); y += 4;
161   }
162   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
163     const float32x4_t vx = vld1q_f32(x); x += 4;
164 
165     const float32x4_t vz = vabsq_f32(vx);
166 
167     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
168     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
169 
170     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
171     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
172     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
173     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
174     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
175     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
176     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
177     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
178 
179     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
180     vn = vsubq_f32(vn, vmagic_bias);
181     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
182 
183     const float32x4_t vp = vmulq_f32(vt, vc1);
184 
185     const float32x4_t vy = vfmaq_f32(vs, vs, vp);
186     const float32x4_t vd = vaddq_f32(vy, vone);
187 
188     float32x4_t vr = vrecpeq_f32(vd);
189     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
190     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
191 
192     float32x4_t vf = vmulq_f32(vy, vr);
193     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
194     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
195     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
196 
197     vst1q_f32(y, vf); y += 4;
198   }
199   if XNN_UNLIKELY(n != 0) {
200     const float32x4_t vx = vld1q_f32(x);
201 
202     const float32x4_t vz = vabsq_f32(vx);
203 
204     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
205     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
206 
207     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
208     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
209     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
210     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
211     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
212     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
213     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
214     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
215 
216     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
217     vn = vsubq_f32(vn, vmagic_bias);
218     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
219 
220     const float32x4_t vp = vmulq_f32(vt, vc1);
221 
222     const float32x4_t vy = vfmaq_f32(vs, vs, vp);
223     const float32x4_t vd = vaddq_f32(vy, vone);
224 
225     float32x4_t vr = vrecpeq_f32(vd);
226     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
227     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
228 
229     float32x4_t vf = vmulq_f32(vy, vr);
230     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
231     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
232     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
233 
234     float32x2_t vf_lo = vget_low_f32(vf);
235     if (n & (2 * sizeof(float))) {
236       vst1_f32(y, vf_lo); y += 2;
237       vf_lo = vget_high_f32(vf);
238     }
239     if (n & (1 * sizeof(float))) {
240       vst1_lane_f32(y, vf_lo, 0);
241     }
242   }
243 }
244