• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/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_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24(
21     size_t n,
22     const float* x,
23     float* y,
24     const void* params) XNN_DISABLE_TSAN
25 {
26   assert(n % sizeof(float) == 0);
27 
28   const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p17f);
29   const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p0f);
30   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31   const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
32   const float32x4_t vc2 = vmovq_n_f32(0x1.FFFF0Ap-2f);
33   const float32x4_t vone = vmovq_n_f32(1.0f);
34   const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
35 
36   for (; n >= 24 * sizeof(float); n -= 24 * 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     const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
42     const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
43 
44     const float32x4_t vz0123 = vabsq_f32(vx0123);
45     const float32x4_t vz4567 = vabsq_f32(vx4567);
46     const float32x4_t vz89AB = vabsq_f32(vx89AB);
47     const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
48     const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
49     const float32x4_t vzKLMN = vabsq_f32(vxKLMN);
50 
51     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
52     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
53     float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
54     float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
55     float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
56     float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e);
57 
58     const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
59     const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
60     const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
61     const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
62     const int32x4_t veGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 17);
63     const int32x4_t veKLMN = vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 17);
64 
65     // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
66     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
67     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
68     const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
69     const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
70     const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask));
71     const uint64x2_t vidxKLMN = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnKLMN), vindex_mask));
72 
73     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
74     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
75     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
76     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
77     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
78     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
79     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
80     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
81     const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
82     const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
83     float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
84     float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
85     const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
86     const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
87     float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
88     float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
89     const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
90     const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
91     float32x2_t vlGH = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxGH]);
92     float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]);
93     const uint64_t vidxKL = vgetq_lane_u64(vidxKLMN, 0);
94     const uint64_t vidxMN = vgetq_lane_u64(vidxKLMN, 1);
95     float32x2_t vlKL = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxKL]);
96     float32x2_t vlMN = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxMN]);
97 
98     vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
99     vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
100     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
101     vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
102     vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
103     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
104     vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
105     vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
106     const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
107     vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
108     vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
109     const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
110     vlGH = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxGH >> 32)], vlGH, 1);
111     vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1);
112     const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ);
113     vlKL = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxKL >> 32)], vlKL, 1);
114     vlMN = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxMN >> 32)], vlMN, 1);
115     const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN);
116 
117     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
118     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
119     const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
120     const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
121     const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ));
122     const float32x4_t vsKLMN = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlKLMN), veKLMN));
123 
124     vn0123 = vsubq_f32(vn0123, vmagic_bias);
125     vn4567 = vsubq_f32(vn4567, vmagic_bias);
126     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
127     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
128     vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
129     vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
130 
131     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
132     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
133     float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
134     float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
135     float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2);
136     float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2);
137 
138     float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
139     float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
140     float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
141     float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
142     float32x4_t vpGHIJ = vmulq_f32(vtGHIJ, vc2);
143     float32x4_t vpKLMN = vmulq_f32(vtKLMN, vc2);
144 
145     vp0123 = vfmsq_f32(vt0123, vp0123, vt0123);
146     vp4567 = vfmsq_f32(vt4567, vp4567, vt4567);
147     vp89AB = vfmsq_f32(vt89AB, vp89AB, vt89AB);
148     vpCDEF = vfmsq_f32(vtCDEF, vpCDEF, vtCDEF);
149     vpGHIJ = vfmsq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
150     vpKLMN = vfmsq_f32(vtKLMN, vpKLMN, vtKLMN);
151 
152     const float32x4_t vy0123 = vfmsq_f32(vs0123, vs0123, vp0123);
153     const float32x4_t vy4567 = vfmsq_f32(vs4567, vs4567, vp4567);
154     const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB);
155     const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF);
156     const float32x4_t vyGHIJ = vfmsq_f32(vsGHIJ, vsGHIJ, vpGHIJ);
157     const float32x4_t vyKLMN = vfmsq_f32(vsKLMN, vsKLMN, vpKLMN);
158 
159     const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
160     const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
161     const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
162     const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
163     const float32x4_t vdGHIJ = vaddq_f32(vyGHIJ, vone);
164     const float32x4_t vdKLMN = vaddq_f32(vyKLMN, vone);
165 
166     float32x4_t vf0123 = vdivq_f32(vy0123, vd0123);
167     float32x4_t vf4567 = vdivq_f32(vy4567, vd4567);
168     float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB);
169     float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF);
170     float32x4_t vfGHIJ = vdivq_f32(vyGHIJ, vdGHIJ);
171     float32x4_t vfKLMN = vdivq_f32(vyKLMN, vdKLMN);
172 
173     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
174     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
175     vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
176     vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
177     vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
178     vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm_cutoff)));
179 
180     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
181     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
182     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
183     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
184     const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
185     const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
186 
187     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
188     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
189     vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
190     vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
191     vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
192     vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN));
193 
194     vst1q_f32(y, vf0123); y += 4;
195     vst1q_f32(y, vf4567); y += 4;
196     vst1q_f32(y, vf89AB); y += 4;
197     vst1q_f32(y, vfCDEF); y += 4;
198     vst1q_f32(y, vfGHIJ); y += 4;
199     vst1q_f32(y, vfKLMN); y += 4;
200   }
201   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
202     const float32x4_t vx = vld1q_f32(x); x += 4;
203 
204     const float32x4_t vz = vabsq_f32(vx);
205 
206     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
207     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
208 
209     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
210     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
211     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
212     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
213     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
214     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
215     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
216     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
217 
218     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
219     vn = vsubq_f32(vn, vmagic_bias);
220     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
221 
222     float32x4_t vp = vmulq_f32(vt, vc2);
223     vp = vfmsq_f32(vt, vp, vt);
224 
225     const float32x4_t vy = vfmsq_f32(vs, vs, vp);
226     const float32x4_t vd = vaddq_f32(vy, vone);
227 
228     float32x4_t vf = vdivq_f32(vy, vd);
229     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
230     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
231     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
232 
233     vst1q_f32(y, vf); y += 4;
234   }
235   if XNN_UNLIKELY(n != 0) {
236     const float32x4_t vx = vld1q_f32(x);
237 
238     const float32x4_t vz = vabsq_f32(vx);
239 
240     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
241     const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
242 
243     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
244     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
245     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
246     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
247     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
248     vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
249     vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
250     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
251 
252     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
253     vn = vsubq_f32(vn, vmagic_bias);
254     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
255 
256     float32x4_t vp = vmulq_f32(vt, vc2);
257     vp = vfmsq_f32(vt, vp, vt);
258 
259     const float32x4_t vy = vfmsq_f32(vs, vs, vp);
260     const float32x4_t vd = vaddq_f32(vy, vone);
261 
262     float32x4_t vf = vdivq_f32(vy, vd);
263     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
264     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
265     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
266 
267     float32x2_t vf_lo = vget_low_f32(vf);
268     if (n & (2 * sizeof(float))) {
269       vst1_f32(y, vf_lo); y += 2;
270       vf_lo = vget_high_f32(vf);
271     }
272     if (n & (1 * sizeof(float))) {
273       vst1_lane_f32(y, vf_lo, 0);
274     }
275   }
276 }
277