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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 float32x4_t vx0123 = vld1q_f32(x); x += 4;
47 float32x4_t vx4567 = vld1q_f32(x); x += 4;
48
49 const float32x4_t vz0123 = vmaxq_f32(vmulq_f32(vx0123, vprescale), vsat_cutoff);
50 const float32x4_t vz4567 = vmaxq_f32(vmulq_f32(vx4567, vprescale), vsat_cutoff);
51
52 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
53 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
54
55 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
56 vn0123 = vsubq_f32(vn0123, vmagic_bias);
57 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
58 vn4567 = vsubq_f32(vn4567, vmagic_bias);
59
60 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
61 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
62
63 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
64 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
65
66 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
67 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
68
69 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
70 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
71
72 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
73 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
74
75 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
76 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
77
78 vp0123 = vmulq_f32(vp0123, vt0123);
79 vp4567 = vmulq_f32(vp4567, vt4567);
80
81 vt0123 = vmulq_f32(vt0123, vs0123);
82 vs0123 = vsubq_f32(vs0123, vone);
83 vt4567 = vmulq_f32(vt4567, vs4567);
84 vs4567 = vsubq_f32(vs4567, vone);
85
86 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
87 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
88
89 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
90 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
91
92 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
93 vx0123 = vmulq_f32(vx0123, vbeta);
94 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
95 vx4567 = vmulq_f32(vx4567, vbeta);
96
97 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
98 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
99
100 vst1q_f32(y, vy0123); y += 4;
101 vst1q_f32(y, vy4567); y += 4;
102 }
103 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
104 float32x4_t vx = vld1q_f32(x); x += 4;
105
106 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
107
108 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
109 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
110 vn = vsubq_f32(vn, vmagic_bias);
111
112 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
113 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
114
115 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
116 vp = vmlaq_f32(vc4, vp, vt);
117 vp = vmlaq_f32(vc3, vp, vt);
118 vp = vmlaq_f32(vc2, vp, vt);
119 vp = vmulq_f32(vp, vt);
120
121 vt = vmulq_f32(vt, vs);
122 vs = vsubq_f32(vs, vone);
123 vp = vmlaq_f32(vt, vp, vt);
124 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
125
126 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
127 vx = vmulq_f32(vx, vbeta);
128 const float32x4_t vy = vbslq_f32(vm, ve, vx);
129
130 vst1q_f32(y, vy); y += 4;
131 }
132 if XNN_UNLIKELY(n != 0) {
133 float32x4_t vx = vld1q_f32(x);
134
135 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
136
137 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
138 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
139 vn = vsubq_f32(vn, vmagic_bias);
140
141 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
142 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
143
144 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
145 vp = vmlaq_f32(vc4, vp, vt);
146 vp = vmlaq_f32(vc3, vp, vt);
147 vp = vmlaq_f32(vc2, vp, vt);
148 vp = vmulq_f32(vp, vt);
149
150 vt = vmulq_f32(vt, vs);
151 vs = vsubq_f32(vs, vone);
152 vp = vmlaq_f32(vt, vp, vt);
153 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
154
155 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
156 vx = vmulq_f32(vx, vbeta);
157 const float32x4_t vy = vbslq_f32(vm, ve, vx);
158
159 float32x2_t vy_lo = vget_low_f32(vy);
160 if (n & (2 * sizeof(float))) {
161 vst1_f32(y, vy_lo); y += 2;
162 vy_lo = vget_high_f32(vy);
163 }
164 if (n & (1 * sizeof(float))) {
165 vst1_lane_f32(y, vy_lo, 0);
166 }
167 }
168 }
169