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__neonfma_rr1_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__neonfma_rr1_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 = vmovq_n_f32(-0x1.62E430p-1f);
37 const float32x4_t vc6 = vmovq_n_f32(0x1.6b7338p-10f);
38 const float32x4_t vc5 = vmovq_n_f32(0x1.12278Ep-7f);
39 const float32x4_t vc4 = vmovq_n_f32(0x1.555716p-5f);
40 const float32x4_t vc3 = vmovq_n_f32(0x1.5554B0p-3f);
41 const float32x4_t vc2 = vmovq_n_f32(0x1.FFFFFEp-2f);
42 const float32x4_t vone = vmovq_n_f32(1.0f);
43
44 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
45 float32x4_t vx0123 = vld1q_f32(x); x += 4;
46 float32x4_t vx4567 = vld1q_f32(x); x += 4;
47 float32x4_t vx89AB = vld1q_f32(x); x += 4;
48 float32x4_t vxCDEF = vld1q_f32(x); x += 4;
49 float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
50
51 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
52 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
53 const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
54 const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
55 const float32x4_t vzGHIJ = vmaxq_f32(vmulq_f32(vxGHIJ, vprescale), vsat_cutoff);
56
57 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vlog2e);
58 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vlog2e);
59 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vlog2e);
60 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vlog2e);
61 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vlog2e);
62
63 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
64 vn0123 = vsubq_f32(vn0123, vmagic_bias);
65 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
66 vn4567 = vsubq_f32(vn4567, vmagic_bias);
67 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
68 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
69 float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
70 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
71 float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
72 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
73
74 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vminus_ln2);
75 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vminus_ln2);
76 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vminus_ln2);
77 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vminus_ln2);
78 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vminus_ln2);
79
80 float32x4_t vp0123 = vfmaq_f32(vc5, vc6, vt0123);
81 float32x4_t vp4567 = vfmaq_f32(vc5, vc6, vt4567);
82 float32x4_t vp89AB = vfmaq_f32(vc5, vc6, vt89AB);
83 float32x4_t vpCDEF = vfmaq_f32(vc5, vc6, vtCDEF);
84 float32x4_t vpGHIJ = vfmaq_f32(vc5, vc6, vtGHIJ);
85
86 vp0123 = vfmaq_f32(vc4, vp0123, vt0123);
87 vp4567 = vfmaq_f32(vc4, vp4567, vt4567);
88 vp89AB = vfmaq_f32(vc4, vp89AB, vt89AB);
89 vpCDEF = vfmaq_f32(vc4, vpCDEF, vtCDEF);
90 vpGHIJ = vfmaq_f32(vc4, vpGHIJ, vtGHIJ);
91
92 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
93 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
94 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
95 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF);
96 vpGHIJ = vfmaq_f32(vc3, vpGHIJ, vtGHIJ);
97
98 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
99 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
100 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
101 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF);
102 vpGHIJ = vfmaq_f32(vc2, vpGHIJ, vtGHIJ);
103
104 vp0123 = vmulq_f32(vp0123, vt0123);
105 vp4567 = vmulq_f32(vp4567, vt4567);
106 vp89AB = vmulq_f32(vp89AB, vt89AB);
107 vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
108 vpGHIJ = vmulq_f32(vpGHIJ, vtGHIJ);
109
110 vt0123 = vmulq_f32(vt0123, vs0123);
111 vs0123 = vsubq_f32(vs0123, vone);
112 vt4567 = vmulq_f32(vt4567, vs4567);
113 vs4567 = vsubq_f32(vs4567, vone);
114 vt89AB = vmulq_f32(vt89AB, vs89AB);
115 vs89AB = vsubq_f32(vs89AB, vone);
116 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
117 vsCDEF = vsubq_f32(vsCDEF, vone);
118 vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
119 vsGHIJ = vsubq_f32(vsGHIJ, vone);
120
121 vp0123 = vfmaq_f32(vt0123, vp0123, vt0123);
122 vp4567 = vfmaq_f32(vt4567, vp4567, vt4567);
123 vp89AB = vfmaq_f32(vt89AB, vp89AB, vt89AB);
124 vpCDEF = vfmaq_f32(vtCDEF, vpCDEF, vtCDEF);
125 vpGHIJ = vfmaq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
126
127 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
128 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
129 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
130 const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
131 const float32x4_t veGHIJ = vmulq_f32(vaddq_f32(vpGHIJ, vsGHIJ), valpha);
132
133 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
134 vx0123 = vmulq_f32(vx0123, vbeta);
135 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
136 vx4567 = vmulq_f32(vx4567, vbeta);
137 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
138 vx89AB = vmulq_f32(vx89AB, vbeta);
139 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
140 vxCDEF = vmulq_f32(vxCDEF, vbeta);
141 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
142 vxGHIJ = vmulq_f32(vxGHIJ, vbeta);
143
144 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
145 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
146 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
147 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
148 const float32x4_t vyGHIJ = vbslq_f32(vmGHIJ, veGHIJ, vxGHIJ);
149
150 vst1q_f32(y, vy0123); y += 4;
151 vst1q_f32(y, vy4567); y += 4;
152 vst1q_f32(y, vy89AB); y += 4;
153 vst1q_f32(y, vyCDEF); y += 4;
154 vst1q_f32(y, vyGHIJ); y += 4;
155 }
156 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
157 float32x4_t vx = vld1q_f32(x); x += 4;
158
159 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
160
161 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
162 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
163 vn = vsubq_f32(vn, vmagic_bias);
164
165 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
166
167 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
168 vp = vfmaq_f32(vc4, vp, vt);
169 vp = vfmaq_f32(vc3, vp, vt);
170 vp = vfmaq_f32(vc2, vp, vt);
171 vp = vmulq_f32(vp, vt);
172
173 vt = vmulq_f32(vt, vs);
174 vs = vsubq_f32(vs, vone);
175 vp = vfmaq_f32(vt, vp, vt);
176 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
177
178 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
179 vx = vmulq_f32(vx, vbeta);
180 const float32x4_t vy = vbslq_f32(vm, ve, vx);
181
182 vst1q_f32(y, vy); y += 4;
183 }
184 if XNN_UNLIKELY(n != 0) {
185 float32x4_t vx = vld1q_f32(x);
186
187 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
188
189 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vlog2e);
190 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
191 vn = vsubq_f32(vn, vmagic_bias);
192
193 float32x4_t vt = vfmaq_f32(vz, vn, vminus_ln2);
194
195 float32x4_t vp = vfmaq_f32(vc5, vc6, vt);
196 vp = vfmaq_f32(vc4, vp, vt);
197 vp = vfmaq_f32(vc3, vp, vt);
198 vp = vfmaq_f32(vc2, vp, vt);
199 vp = vmulq_f32(vp, vt);
200
201 vt = vmulq_f32(vt, vs);
202 vs = vsubq_f32(vs, vone);
203 vp = vfmaq_f32(vt, vp, vt);
204 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
205
206 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
207 vx = vmulq_f32(vx, vbeta);
208 const float32x4_t vy = vbslq_f32(vm, ve, vx);
209
210 float32x2_t vy_lo = vget_low_f32(vy);
211 if (n & (2 * sizeof(float))) {
212 vst1_f32(y, vy_lo); y += 2;
213 vy_lo = vget_high_f32(vy);
214 }
215 if (n & (1 * sizeof(float))) {
216 vst1_lane_f32(y, vy_lo, 0);
217 }
218 }
219 }
220