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_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 >= 16 * sizeof(float); n -= 16 * 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
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
56 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
57 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
58 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
59 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vlog2e);
60
61 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
62 vn0123 = vsubq_f32(vn0123, vmagic_bias);
63 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
64 vn4567 = vsubq_f32(vn4567, vmagic_bias);
65 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
66 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
67 float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
68 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
69
70 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
71 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
72 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
73 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vminus_ln2_hi);
74
75 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
76 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
77 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
78 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);
79
80 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
81 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
82 float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
83 float32x4_t vpCDEF = vmlaq_f32(vc5, vc6, vtCDEF);
84
85 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
86 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
87 vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
88 vpCDEF = vmlaq_f32(vc4, vpCDEF, vtCDEF);
89
90 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
91 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
92 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
93 vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
94
95 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
96 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
97 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
98 vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
99
100 vp0123 = vmulq_f32(vp0123, vt0123);
101 vp4567 = vmulq_f32(vp4567, vt4567);
102 vp89AB = vmulq_f32(vp89AB, vt89AB);
103 vpCDEF = vmulq_f32(vpCDEF, vtCDEF);
104
105 vt0123 = vmulq_f32(vt0123, vs0123);
106 vs0123 = vsubq_f32(vs0123, vone);
107 vt4567 = vmulq_f32(vt4567, vs4567);
108 vs4567 = vsubq_f32(vs4567, vone);
109 vt89AB = vmulq_f32(vt89AB, vs89AB);
110 vs89AB = vsubq_f32(vs89AB, vone);
111 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
112 vsCDEF = vsubq_f32(vsCDEF, vone);
113
114 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
115 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
116 vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
117 vpCDEF = vmlaq_f32(vtCDEF, vpCDEF, vtCDEF);
118
119 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
120 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
121 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
122 const float32x4_t veCDEF = vmulq_f32(vaddq_f32(vpCDEF, vsCDEF), valpha);
123
124 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
125 vx0123 = vmulq_f32(vx0123, vbeta);
126 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
127 vx4567 = vmulq_f32(vx4567, vbeta);
128 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
129 vx89AB = vmulq_f32(vx89AB, vbeta);
130 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
131 vxCDEF = vmulq_f32(vxCDEF, vbeta);
132
133 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
134 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
135 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
136 const float32x4_t vyCDEF = vbslq_f32(vmCDEF, veCDEF, vxCDEF);
137
138 vst1q_f32(y, vy0123); y += 4;
139 vst1q_f32(y, vy4567); y += 4;
140 vst1q_f32(y, vy89AB); y += 4;
141 vst1q_f32(y, vyCDEF); y += 4;
142 }
143 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
144 float32x4_t vx = vld1q_f32(x); x += 4;
145
146 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
147
148 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
149 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
150 vn = vsubq_f32(vn, vmagic_bias);
151
152 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
153 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
154
155 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
156 vp = vmlaq_f32(vc4, vp, vt);
157 vp = vmlaq_f32(vc3, vp, vt);
158 vp = vmlaq_f32(vc2, vp, vt);
159 vp = vmulq_f32(vp, vt);
160
161 vt = vmulq_f32(vt, vs);
162 vs = vsubq_f32(vs, vone);
163 vp = vmlaq_f32(vt, vp, vt);
164 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
165
166 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
167 vx = vmulq_f32(vx, vbeta);
168 const float32x4_t vy = vbslq_f32(vm, ve, vx);
169
170 vst1q_f32(y, vy); y += 4;
171 }
172 if XNN_UNLIKELY(n != 0) {
173 float32x4_t vx = vld1q_f32(x);
174
175 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
176
177 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
178 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
179 vn = vsubq_f32(vn, vmagic_bias);
180
181 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
182 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
183
184 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
185 vp = vmlaq_f32(vc4, vp, vt);
186 vp = vmlaq_f32(vc3, vp, vt);
187 vp = vmlaq_f32(vc2, vp, vt);
188 vp = vmulq_f32(vp, vt);
189
190 vt = vmulq_f32(vt, vs);
191 vs = vsubq_f32(vs, vone);
192 vp = vmlaq_f32(vt, vp, vt);
193 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
194
195 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
196 vx = vmulq_f32(vx, vbeta);
197 const float32x4_t vy = vbslq_f32(vm, ve, vx);
198
199 float32x2_t vy_lo = vget_low_f32(vy);
200 if (n & (2 * sizeof(float))) {
201 vst1_f32(y, vy_lo); y += 2;
202 vy_lo = vget_high_f32(vy);
203 }
204 if (n & (1 * sizeof(float))) {
205 vst1_lane_f32(y, vy_lo, 0);
206 }
207 }
208 }
209