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_x12(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_x12(
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 >= 12 * sizeof(float); n -= 12 * 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
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
54 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vlog2e);
55 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vlog2e);
56 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vlog2e);
57
58 float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
59 vn0123 = vsubq_f32(vn0123, vmagic_bias);
60 float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
61 vn4567 = vsubq_f32(vn4567, vmagic_bias);
62 float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
63 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
64
65 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vminus_ln2_hi);
66 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vminus_ln2_hi);
67 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vminus_ln2_hi);
68
69 vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);
70 vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);
71 vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);
72
73 float32x4_t vp0123 = vmlaq_f32(vc5, vc6, vt0123);
74 float32x4_t vp4567 = vmlaq_f32(vc5, vc6, vt4567);
75 float32x4_t vp89AB = vmlaq_f32(vc5, vc6, vt89AB);
76
77 vp0123 = vmlaq_f32(vc4, vp0123, vt0123);
78 vp4567 = vmlaq_f32(vc4, vp4567, vt4567);
79 vp89AB = vmlaq_f32(vc4, vp89AB, vt89AB);
80
81 vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
82 vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
83 vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
84
85 vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
86 vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
87 vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
88
89 vp0123 = vmulq_f32(vp0123, vt0123);
90 vp4567 = vmulq_f32(vp4567, vt4567);
91 vp89AB = vmulq_f32(vp89AB, vt89AB);
92
93 vt0123 = vmulq_f32(vt0123, vs0123);
94 vs0123 = vsubq_f32(vs0123, vone);
95 vt4567 = vmulq_f32(vt4567, vs4567);
96 vs4567 = vsubq_f32(vs4567, vone);
97 vt89AB = vmulq_f32(vt89AB, vs89AB);
98 vs89AB = vsubq_f32(vs89AB, vone);
99
100 vp0123 = vmlaq_f32(vt0123, vp0123, vt0123);
101 vp4567 = vmlaq_f32(vt4567, vp4567, vt4567);
102 vp89AB = vmlaq_f32(vt89AB, vp89AB, vt89AB);
103
104 const float32x4_t ve0123 = vmulq_f32(vaddq_f32(vp0123, vs0123), valpha);
105 const float32x4_t ve4567 = vmulq_f32(vaddq_f32(vp4567, vs4567), valpha);
106 const float32x4_t ve89AB = vmulq_f32(vaddq_f32(vp89AB, vs89AB), valpha);
107
108 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
109 vx0123 = vmulq_f32(vx0123, vbeta);
110 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
111 vx4567 = vmulq_f32(vx4567, vbeta);
112 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
113 vx89AB = vmulq_f32(vx89AB, vbeta);
114
115 const float32x4_t vy0123 = vbslq_f32(vm0123, ve0123, vx0123);
116 const float32x4_t vy4567 = vbslq_f32(vm4567, ve4567, vx4567);
117 const float32x4_t vy89AB = vbslq_f32(vm89AB, ve89AB, vx89AB);
118
119 vst1q_f32(y, vy0123); y += 4;
120 vst1q_f32(y, vy4567); y += 4;
121 vst1q_f32(y, vy89AB); y += 4;
122 }
123 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
124 float32x4_t vx = vld1q_f32(x); x += 4;
125
126 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
127
128 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
129 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
130 vn = vsubq_f32(vn, vmagic_bias);
131
132 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
133 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
134
135 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
136 vp = vmlaq_f32(vc4, vp, vt);
137 vp = vmlaq_f32(vc3, vp, vt);
138 vp = vmlaq_f32(vc2, vp, vt);
139 vp = vmulq_f32(vp, vt);
140
141 vt = vmulq_f32(vt, vs);
142 vs = vsubq_f32(vs, vone);
143 vp = vmlaq_f32(vt, vp, vt);
144 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
145
146 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
147 vx = vmulq_f32(vx, vbeta);
148 const float32x4_t vy = vbslq_f32(vm, ve, vx);
149
150 vst1q_f32(y, vy); y += 4;
151 }
152 if XNN_UNLIKELY(n != 0) {
153 float32x4_t vx = vld1q_f32(x);
154
155 const float32x4_t vz = vmaxq_f32(vmulq_f32(vx, vprescale), vsat_cutoff);
156
157 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vlog2e);
158 float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
159 vn = vsubq_f32(vn, vmagic_bias);
160
161 float32x4_t vt = vmlaq_f32(vz, vn, vminus_ln2_hi);
162 vt = vmlaq_f32(vt, vn, vminus_ln2_lo);
163
164 float32x4_t vp = vmlaq_f32(vc5, vc6, vt);
165 vp = vmlaq_f32(vc4, vp, vt);
166 vp = vmlaq_f32(vc3, vp, vt);
167 vp = vmlaq_f32(vc2, vp, vt);
168 vp = vmulq_f32(vp, vt);
169
170 vt = vmulq_f32(vt, vs);
171 vs = vsubq_f32(vs, vone);
172 vp = vmlaq_f32(vt, vp, vt);
173 const float32x4_t ve = vmulq_f32(vaddq_f32(vp, vs), valpha);
174
175 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
176 vx = vmulq_f32(vx, vbeta);
177 const float32x4_t vy = vbslq_f32(vm, ve, vx);
178
179 float32x2_t vy_lo = vget_low_f32(vy);
180 if (n & (2 * sizeof(float))) {
181 vst1_f32(y, vy_lo); y += 2;
182 vy_lo = vget_high_f32(vy);
183 }
184 if (n & (1 * sizeof(float))) {
185 vst1_lane_f32(y, vy_lo, 0);
186 }
187 }
188 }
189