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_x16(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_x16(
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_OOB_READS
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->neon_rr2_p6.prescale);
30 const float32x4_t valpha = vld1q_dup_f32(¶ms->neon_rr2_p6.alpha);
31 const float32x4_t vbeta = vld1q_dup_f32(¶ms->neon_rr2_p6.beta);
32 const float32x4_t vsat_cutoff = vld1q_dup_f32(¶ms->neon_rr2_p6.sat_cutoff);
33 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon_rr2_p6.magic_bias);
34 const float32x4_t vlog2e = vld1q_dup_f32(¶ms->neon_rr2_p6.log2e);
35 const float32x4_t vminus_ln2_hi = vld1q_dup_f32(¶ms->neon_rr2_p6.minus_ln2_hi);
36 const float32x4_t vminus_ln2_lo = vld1q_dup_f32(¶ms->neon_rr2_p6.minus_ln2_lo);
37 const float32x4_t vc6 = vld1q_dup_f32(¶ms->neon_rr2_p6.c6);
38 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neon_rr2_p6.c5);
39 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neon_rr2_p6.c4);
40 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neon_rr2_p6.c3);
41 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neon_rr2_p6.c2);
42 const float32x4_t vone = vmovq_n_f32(1.0f);
43
44 for (; n >= 16 * sizeof(float); n -= 16 * 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
50 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
51 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
52 const float32x4_t vz89AB = vmaxq_f32(vmulq_f32(vx89AB, vprescale), vsat_cutoff);
53 const float32x4_t vzCDEF = vmaxq_f32(vmulq_f32(vxCDEF, vprescale), vsat_cutoff);
54
55 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
56 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
57 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
58 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
59
60 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
61 vn0123 = vsubq_f32(vn0123, vmagic_bias);
62 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
63 vn4567 = vsubq_f32(vn4567, vmagic_bias);
64 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
65 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
66 float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
67 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
68
69 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
70 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
71 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
72 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
73
74 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
75 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
76 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
77 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
78
79 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
80 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
81 float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
82 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF);
83
84 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
85 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
86 vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
87 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF);
88
89 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
90 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
91 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
92 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
93
94 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
95 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
96 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
97 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
98
99 vp0123 = vmulq_f32(vp0123, vt0123);
100 vp4567 = vmulq_f32(vp4567, vt4567);
101 vp89AB = vmulq_f32(vp89AB, vt89AB);
102 vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
103
104 vt0123 = vmulq_f32(vt0123, vs0123);
105 vs0123 = vsubq_f32(vs0123, vone);
106 vt4567 = vmulq_f32(vt4567, vs4567);
107 vs4567 = vsubq_f32(vs4567, vone);
108 vt89AB = vmulq_f32(vt89AB, vs89AB);
109 vs89AB = vsubq_f32(vs89AB, vone);
110 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
111 vsCDEF = vsubq_f32(vsCDEF, vone);
112
113 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
114 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
115 vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
116 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
117
118 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
119 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
120 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
121 const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
122
123 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
124 vx0123 = vmulq_f32(vx0123, vbeta);
125 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
126 vx4567 = vmulq_f32(vx4567, vbeta);
127 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
128 vx89AB = vmulq_f32(vx89AB, vbeta);
129 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
130 vxCDEF = vmulq_f32(vxCDEF, vbeta);
131
132 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
133 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
134 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
135 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
136
137 vst1q_f32(y, vy0123); y += 4;
138 vst1q_f32(y, vy4567); y += 4;
139 vst1q_f32(y, vy89AB); y += 4;
140 vst1q_f32(y, vyCDEF); y += 4;
141 }
142 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
143 float32x4_t vx = vld1q_f32(x); x += 4;
144
145 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
146
147 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
148 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
149 vn = vsubq_f32(vn, vmagic_bias);
150
151 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
152 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
153
154 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
155 vp = vmlaq_f32(vc4, vp, vt);
156 vp = vmlaq_f32(vc3, vp, vt);
157 vp = vmlaq_f32(vc2, vp, vt);
158 vp = vmulq_f32(vp, vt);
159
160 vt = vmulq_f32(vt, vs);
161 vs = vsubq_f32(vs, vone);
162 vp = vmlaq_f32(vt, vp, vt);
163 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
164
165 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
166 vx = vmulq_f32(vx, vbeta);
167 const float32x4_t vy = vbslq_f32(vm, ve, vx);
168
169 vst1q_f32(y, vy); y += 4;
170 }
171 if XNN_UNLIKELY(n != 0) {
172 float32x4_t vx = vld1q_f32(x);
173
174 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
175
176 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
177 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
178 vn = vsubq_f32(vn, vmagic_bias);
179
180 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
181 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
182
183 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
184 vp = vmlaq_f32(vc4, vp, vt);
185 vp = vmlaq_f32(vc3, vp, vt);
186 vp = vmlaq_f32(vc2, vp, vt);
187 vp = vmulq_f32(vp, vt);
188
189 vt = vmulq_f32(vt, vs);
190 vs = vsubq_f32(vs, vone);
191 vp = vmlaq_f32(vt, vp, vt);
192 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
193
194 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
195 vx = vmulq_f32(vx, vbeta);
196 const float32x4_t vy = vbslq_f32(vm, ve, vx);
197
198 float32x2_t vy_lo = vget_low_f32(vy);
199 if (n & (2 * sizeof(float))) {
200 vst1_f32(y, vy_lo); y += 2;
201 vy_lo = vget_high_f32(vy);
202 }
203 if (n & (1 * sizeof(float))) {
204 vst1_lane_f32(y, vy_lo, 0);
205 }
206 }
207 }
208