1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/neon-p6.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__neon_rr2_p6_x20(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__neon_rr2_p6_x20(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const float32x4_t vprescale = vld1q_dup_f32(¶ms->scalar.prescale);
30 const float32x4_t valpha = vld1q_dup_f32(¶ms->scalar.alpha);
31 const float32x4_t vbeta = vld1q_dup_f32(¶ms->scalar.beta);
32
33 const float32x4_t vsat_cutoff = vmovq_n_f32(-0x1.154246p+4f);
34 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.8000FEp23f);
35 const float32x4_t vlog2e = vmovq_n_f32(0x1.715476p+0f);
36 const float32x4_t vminus_ln2_hi = vmovq_n_f32(-0x1.62E440p-1f);
37 const float32x4_t vminus_ln2_lo = vmovq_n_f32(0x1.0105C6p-21f);
38 const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
39 const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
40 const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
41 const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
42 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
43 const float32x4_t vone = vmovq_n_f32(1.0f);
44
45 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
46 float32x4_t vx0123 = vld1q_f32(x); x += 4;
47 float32x4_t vx4567 = vld1q_f32(x); x += 4;
48 float32x4_t vx89AB = vld1q_f32(x); x += 4;
49 float32x4_t vxCDEF = vld1q_f32(x); x += 4;
50 float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
51
52 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
53 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
54 const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
55 const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
56 const float32x4_t vzGHIJ = vmaxq_f32(vmulq_f32(vxGHIJ, vprescale), vsat_cutoff);
57
58 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
59 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
60 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
61 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
62 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vlog2e);
63
64 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
65 vn0123 = vsubq_f32(vn0123, vmagic_bias);
66 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
67 vn4567 = vsubq_f32(vn4567, vmagic_bias);
68 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
69 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
70 float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
71 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
72 float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
73 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
74
75 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
76 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
77 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
78 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
79 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vminus_ln2_hi);
80
81 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
82 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
83 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
84 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
85 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vminus_ln2_lo);
86
87 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
88 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
89 float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
90 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF);
91 float32x4_t vpGHIJ = vmlaq_f32(vc5, vc6, vtGHIJ);
92
93 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
94 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
95 vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
96 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF);
97 vpGHIJ = vmlaq_f32(vc4, vpGHIJ, vtGHIJ);
98
99 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
100 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
101 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
102 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
103 vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);
104
105 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
106 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
107 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
108 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
109 vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);
110
111 vp0123 = vmulq_f32(vp0123, vt0123);
112 vp4567 = vmulq_f32(vp4567, vt4567);
113 vp89AB = vmulq_f32(vp89AB, vt89AB);
114 vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
115 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ);
116
117 vt0123 = vmulq_f32(vt0123, vs0123);
118 vs0123 = vsubq_f32(vs0123, vone);
119 vt4567 = vmulq_f32(vt4567, vs4567);
120 vs4567 = vsubq_f32(vs4567, vone);
121 vt89AB = vmulq_f32(vt89AB, vs89AB);
122 vs89AB = vsubq_f32(vs89AB, vone);
123 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
124 vsCDEF = vsubq_f32(vsCDEF, vone);
125 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
126 vsGHIJ = vsubq_f32(vsGHIJ, vone);
127
128 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
129 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
130 vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
131 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
132 vpGHIJ = vmlaq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
133
134 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
135 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
136 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
137 const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
138 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha);
139
140 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
141 vx0123 = vmulq_f32(vx0123, vbeta);
142 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
143 vx4567 = vmulq_f32(vx4567, vbeta);
144 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
145 vx89AB = vmulq_f32(vx89AB, vbeta);
146 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
147 vxCDEF = vmulq_f32(vxCDEF, vbeta);
148 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
149 vxGHIJ = vmulq_f32(vxGHIJ, vbeta);
150
151 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
152 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
153 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
154 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
155 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ);
156
157 vst1q_f32(y, vy0123); y += 4;
158 vst1q_f32(y, vy4567); y += 4;
159 vst1q_f32(y, vy89AB); y += 4;
160 vst1q_f32(y, vyCDEF); y += 4;
161 vst1q_f32(y, vyGHIJ); y += 4;
162 }
163 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
164 float32x4_t vx = vld1q_f32(x); x += 4;
165
166 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
167
168 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
169 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
170 vn = vsubq_f32(vn, vmagic_bias);
171
172 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
173 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
174
175 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
176 vp = vmlaq_f32(vc4, vp, vt);
177 vp = vmlaq_f32(vc3, vp, vt);
178 vp = vmlaq_f32(vc2, vp, vt);
179 vp = vmulq_f32(vp, vt);
180
181 vt = vmulq_f32(vt, vs);
182 vs = vsubq_f32(vs, vone);
183 vp = vmlaq_f32(vt, vp, vt);
184 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
185
186 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
187 vx = vmulq_f32(vx, vbeta);
188 const float32x4_t vy = vbslq_f32(vm, ve, vx);
189
190 vst1q_f32(y, vy); y += 4;
191 }
192 if XNN_UNLIKELY(n != 0) {
193 float32x4_t vx = vld1q_f32(x);
194
195 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
196
197 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
198 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
199 vn = vsubq_f32(vn, vmagic_bias);
200
201 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
202 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
203
204 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
205 vp = vmlaq_f32(vc4, vp, vt);
206 vp = vmlaq_f32(vc3, vp, vt);
207 vp = vmlaq_f32(vc2, vp, vt);
208 vp = vmulq_f32(vp, vt);
209
210 vt = vmulq_f32(vt, vs);
211 vs = vsubq_f32(vs, vone);
212 vp = vmlaq_f32(vt, vp, vt);
213 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
214
215 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
216 vx = vmulq_f32(vx, vbeta);
217 const float32x4_t vy = vbslq_f32(vm, ve, vx);
218
219 float32x2_t vy_lo = vget_low_f32(vy);
220 if (n & (2 * sizeof(float))) {
221 vst1_f32(y, vy_lo); y += 2;
222 vy_lo = vget_high_f32(vy);
223 }
224 if (n & (1 * sizeof(float))) {
225 vst1_lane_f32(y, vy_lo, 0);
226 }
227 }
228 }
229